1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //=============================================================================
45 #define NODE_HASH_MINIMUM_SIZE 255
46
47 //------------------------------NodeHash---------------------------------------
48 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
49 _a(arena),
50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
51 _inserts(0), _insert_limit( insert_limit() ),
52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
53 #ifndef PRODUCT
54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
56 _total_inserts(0), _total_insert_probes(0)
57 #endif
58 {
59 // _sentinel must be in the current node space
60 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
61 memset(_table,0,sizeof(Node*)*_max);
62 }
63
64 //------------------------------hash_find--------------------------------------
65 // Find in hash table
66 Node *NodeHash::hash_find( const Node *n ) {
67 // ((Node*)n)->set_hash( n->hash() );
68 uint hash = n->hash();
69 if (hash == Node::NO_HASH) {
70 NOT_PRODUCT( _lookup_misses++ );
71 return nullptr;
72 }
73 uint key = hash & (_max-1);
74 uint stride = key | 0x01;
75 NOT_PRODUCT( _look_probes++ );
76 Node *k = _table[key]; // Get hashed value
77 if( !k ) { // ?Miss?
78 NOT_PRODUCT( _lookup_misses++ );
79 return nullptr; // Miss!
80 }
81
82 int op = n->Opcode();
83 uint req = n->req();
84 while( 1 ) { // While probing hash table
85 if( k->req() == req && // Same count of inputs
86 k->Opcode() == op ) { // Same Opcode
87 for( uint i=0; i<req; i++ )
88 if( n->in(i)!=k->in(i)) // Different inputs?
89 goto collision; // "goto" is a speed hack...
90 if( n->cmp(*k) ) { // Check for any special bits
91 NOT_PRODUCT( _lookup_hits++ );
92 return k; // Hit!
93 }
94 }
95 collision:
96 NOT_PRODUCT( _look_probes++ );
97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
98 k = _table[key]; // Get hashed value
99 if( !k ) { // ?Miss?
100 NOT_PRODUCT( _lookup_misses++ );
101 return nullptr; // Miss!
102 }
103 }
104 ShouldNotReachHere();
105 return nullptr;
106 }
107
108 //------------------------------hash_find_insert-------------------------------
109 // Find in hash table, insert if not already present
110 // Used to preserve unique entries in hash table
111 Node *NodeHash::hash_find_insert( Node *n ) {
112 // n->set_hash( );
113 uint hash = n->hash();
114 if (hash == Node::NO_HASH) {
115 NOT_PRODUCT( _lookup_misses++ );
116 return nullptr;
117 }
118 uint key = hash & (_max-1);
119 uint stride = key | 0x01; // stride must be relatively prime to table siz
120 uint first_sentinel = 0; // replace a sentinel if seen.
121 NOT_PRODUCT( _look_probes++ );
122 Node *k = _table[key]; // Get hashed value
123 if( !k ) { // ?Miss?
124 NOT_PRODUCT( _lookup_misses++ );
125 _table[key] = n; // Insert into table!
126 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
127 check_grow(); // Grow table if insert hit limit
128 return nullptr; // Miss!
129 }
130 else if( k == _sentinel ) {
131 first_sentinel = key; // Can insert here
132 }
133
134 int op = n->Opcode();
135 uint req = n->req();
136 while( 1 ) { // While probing hash table
137 if( k->req() == req && // Same count of inputs
138 k->Opcode() == op ) { // Same Opcode
139 for( uint i=0; i<req; i++ )
140 if( n->in(i)!=k->in(i)) // Different inputs?
141 goto collision; // "goto" is a speed hack...
142 if( n->cmp(*k) ) { // Check for any special bits
143 NOT_PRODUCT( _lookup_hits++ );
144 return k; // Hit!
145 }
146 }
147 collision:
148 NOT_PRODUCT( _look_probes++ );
149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
150 k = _table[key]; // Get hashed value
151 if( !k ) { // ?Miss?
152 NOT_PRODUCT( _lookup_misses++ );
153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
154 _table[key] = n; // Insert into table!
155 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
156 check_grow(); // Grow table if insert hit limit
157 return nullptr; // Miss!
158 }
159 else if( first_sentinel == 0 && k == _sentinel ) {
160 first_sentinel = key; // Can insert here
161 }
162
163 }
164 ShouldNotReachHere();
165 return nullptr;
166 }
167
168 //------------------------------hash_insert------------------------------------
169 // Insert into hash table
170 void NodeHash::hash_insert( Node *n ) {
171 // // "conflict" comments -- print nodes that conflict
172 // bool conflict = false;
173 // n->set_hash();
174 uint hash = n->hash();
175 if (hash == Node::NO_HASH) {
176 return;
177 }
178 check_grow();
179 uint key = hash & (_max-1);
180 uint stride = key | 0x01;
181
182 while( 1 ) { // While probing hash table
183 NOT_PRODUCT( _insert_probes++ );
184 Node *k = _table[key]; // Get hashed value
185 if( !k || (k == _sentinel) ) break; // Found a slot
186 assert( k != n, "already inserted" );
187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
189 }
190 _table[key] = n; // Insert into table!
191 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
192 // if( conflict ) { n->dump(); }
193 }
194
195 //------------------------------hash_delete------------------------------------
196 // Replace in hash table with sentinel
197 bool NodeHash::hash_delete( const Node *n ) {
198 Node *k;
199 uint hash = n->hash();
200 if (hash == Node::NO_HASH) {
201 NOT_PRODUCT( _delete_misses++ );
202 return false;
203 }
204 uint key = hash & (_max-1);
205 uint stride = key | 0x01;
206 DEBUG_ONLY( uint counter = 0; );
207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
208 DEBUG_ONLY( counter++ );
209 NOT_PRODUCT( _delete_probes++ );
210 k = _table[key]; // Get hashed value
211 if( !k ) { // Miss?
212 NOT_PRODUCT( _delete_misses++ );
213 return false; // Miss! Not in chain
214 }
215 else if( n == k ) {
216 NOT_PRODUCT( _delete_hits++ );
217 _table[key] = _sentinel; // Hit! Label as deleted entry
218 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
219 return true;
220 }
221 else {
222 // collision: move through table with prime offset
223 key = (key + stride/*7*/) & (_max-1);
224 assert( counter <= _insert_limit, "Cycle in hash-table");
225 }
226 }
227 ShouldNotReachHere();
228 return false;
229 }
230
231 //------------------------------round_up---------------------------------------
232 // Round up to nearest power of 2
233 uint NodeHash::round_up(uint x) {
234 x += (x >> 2); // Add 25% slop
235 return MAX2(16U, round_up_power_of_2(x));
236 }
237
238 //------------------------------grow-------------------------------------------
239 // Grow _table to next power of 2 and insert old entries
240 void NodeHash::grow() {
241 // Record old state
242 uint old_max = _max;
243 Node **old_table = _table;
244 // Construct new table with twice the space
245 #ifndef PRODUCT
246 _grows++;
247 _total_inserts += _inserts;
248 _total_insert_probes += _insert_probes;
249 _insert_probes = 0;
250 #endif
251 _inserts = 0;
252 _max = _max << 1;
253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
254 memset(_table,0,sizeof(Node*)*_max);
255 _insert_limit = insert_limit();
256 // Insert old entries into the new table
257 for( uint i = 0; i < old_max; i++ ) {
258 Node *m = *old_table++;
259 if( !m || m == _sentinel ) continue;
260 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
261 hash_insert(m);
262 }
263 }
264
265 //------------------------------clear------------------------------------------
266 // Clear all entries in _table to null but keep storage
267 void NodeHash::clear() {
268 #ifdef ASSERT
269 // Unlock all nodes upon removal from table.
270 for (uint i = 0; i < _max; i++) {
271 Node* n = _table[i];
272 if (!n || n == _sentinel) continue;
273 n->exit_hash_lock();
274 }
275 #endif
276
277 memset( _table, 0, _max * sizeof(Node*) );
278 }
279
280 //-----------------------remove_useless_nodes----------------------------------
281 // Remove useless nodes from value table,
282 // implementation does not depend on hash function
283 void NodeHash::remove_useless_nodes(VectorSet &useful) {
284
285 // Dead nodes in the hash table inherited from GVN should not replace
286 // existing nodes, remove dead nodes.
287 uint max = size();
288 Node *sentinel_node = sentinel();
289 for( uint i = 0; i < max; ++i ) {
290 Node *n = at(i);
291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
292 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
293 _table[i] = sentinel_node; // Replace with placeholder
294 }
295 }
296 }
297
298
299 void NodeHash::check_no_speculative_types() {
300 #ifdef ASSERT
301 uint max = size();
302 Unique_Node_List live_nodes;
303 Compile::current()->identify_useful_nodes(live_nodes);
304 Node *sentinel_node = sentinel();
305 for (uint i = 0; i < max; ++i) {
306 Node *n = at(i);
307 if (n != nullptr &&
308 n != sentinel_node &&
309 n->is_Type() &&
310 live_nodes.member(n)) {
311 TypeNode* tn = n->as_Type();
312 const Type* t = tn->type();
313 const Type* t_no_spec = t->remove_speculative();
314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
315 }
316 }
317 #endif
318 }
319
320 #ifndef PRODUCT
321 //------------------------------dump-------------------------------------------
322 // Dump statistics for the hash table
323 void NodeHash::dump() {
324 _total_inserts += _inserts;
325 _total_insert_probes += _insert_probes;
326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
327 if (WizardMode) {
328 for (uint i=0; i<_max; i++) {
329 if (_table[i])
330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
331 }
332 }
333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
337 // sentinels increase lookup cost, but not insert cost
338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
339 assert( _inserts+(_inserts>>3) < _max, "table too full" );
340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
341 }
342 }
343
344 Node *NodeHash::find_index(uint idx) { // For debugging
345 // Find an entry by its index value
346 for( uint i = 0; i < _max; i++ ) {
347 Node *m = _table[i];
348 if( !m || m == _sentinel ) continue;
349 if( m->_idx == (uint)idx ) return m;
350 }
351 return nullptr;
352 }
353 #endif
354
355 #ifdef ASSERT
356 NodeHash::~NodeHash() {
357 // Unlock all nodes upon destruction of table.
358 if (_table != (Node**)badAddress) clear();
359 }
360 #endif
361
362
363 //=============================================================================
364 //------------------------------PhaseRemoveUseless-----------------------------
365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
368 // Implementation requires an edge from root to each SafePointNode
369 // at a backward branch. Inserted in add_safepoint().
370
371 // Identify nodes that are reachable from below, useful.
372 C->identify_useful_nodes(_useful);
373 // Update dead node list
374 C->update_dead_node_list(_useful);
375
376 // Remove all useless nodes from PhaseValues' recorded types
377 // Must be done before disconnecting nodes to preserve hash-table-invariant
378 gvn->remove_useless_nodes(_useful.member_set());
379
380 // Remove all useless nodes from future worklist
381 worklist.remove_useless_nodes(_useful.member_set());
382
383 // Disconnect 'useless' nodes that are adjacent to useful nodes
384 C->disconnect_useless_nodes(_useful, worklist);
385 }
386
387 //=============================================================================
388 //------------------------------PhaseRenumberLive------------------------------
389 // First, remove useless nodes (equivalent to identifying live nodes).
390 // Then, renumber live nodes.
391 //
392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
395 // value in the range [0, x).
396 //
397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
399 //
400 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
402 // be processed by IGVN after removal of the useless nodes.
403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
404 // the node's type. The mapping is updated to use the new node IDs as well. We
405 // create a new map, and swap it with the old one.
406 //
407 // Other data structures used by the compiler are not updated. The hash table for value
408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
409 // computing the hash values is not based on node IDs.
410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
411 Unique_Node_List& worklist,
412 PhaseNumber phase_num) :
413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
414 _new_type_array(C->comp_arena()),
415 _old2new_map(C->unique(), C->unique(), -1),
416 _is_pass_finished(false),
417 _live_node_count(C->live_nodes())
418 {
419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
421 assert(_delayed.size() == 0, "should be empty");
422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
423 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
424
425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
426 if (old_node_note_array != nullptr) {
427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
428 new_size = MAX2(8, new_size);
429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
430 C->grow_node_notes(C->node_note_array(), new_size);
431 }
432
433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
434
435 // Iterate over the set of live nodes.
436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
437 Node* n = _useful.at(current_idx);
438
439 const Type* type = gvn->type_or_null(n);
440 _new_type_array.map(current_idx, type);
441
442 assert(_old2new_map.at(n->_idx) == -1, "already seen");
443 _old2new_map.at_put(n->_idx, current_idx);
444
445 if (old_node_note_array != nullptr) {
446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
447 C->set_node_notes_at(current_idx, nn);
448 }
449
450 n->set_idx(current_idx); // Update node ID.
451
452 if (update_embedded_ids(n) < 0) {
453 _delayed.push(n); // has embedded IDs; handle later
454 }
455 }
456
457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
458 worklist.recompute_idx_set();
459
460 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
461
462 _is_pass_finished = true; // pass finished; safe to process delayed updates
463
464 while (_delayed.size() > 0) {
465 Node* n = _delayed.pop();
466 int no_of_updates = update_embedded_ids(n);
467 assert(no_of_updates > 0, "should be updated");
468 }
469
470 // Replace the compiler's type information with the updated type information.
471 gvn->types().swap(_new_type_array);
472
473 // Update the unique node count of the compilation to the number of currently live nodes.
474 C->set_unique(_live_node_count);
475
476 // Set the dead node count to 0 and reset dead node list.
477 C->reset_dead_node_list();
478 }
479
480 int PhaseRenumberLive::new_index(int old_idx) {
481 assert(_is_pass_finished, "not finished");
482 if (_old2new_map.at(old_idx) == -1) { // absent
483 // Allocate a placeholder to preserve uniqueness
484 _old2new_map.at_put(old_idx, _live_node_count);
485 _live_node_count++;
486 }
487 return _old2new_map.at(old_idx);
488 }
489
490 int PhaseRenumberLive::update_embedded_ids(Node* n) {
491 int no_of_updates = 0;
492 if (n->is_Phi()) {
493 PhiNode* phi = n->as_Phi();
494 if (phi->_inst_id != -1) {
495 if (!_is_pass_finished) {
496 return -1; // delay
497 }
498 int new_idx = new_index(phi->_inst_id);
499 assert(new_idx != -1, "");
500 phi->_inst_id = new_idx;
501 no_of_updates++;
502 }
503 if (phi->_inst_mem_id != -1) {
504 if (!_is_pass_finished) {
505 return -1; // delay
506 }
507 int new_idx = new_index(phi->_inst_mem_id);
508 assert(new_idx != -1, "");
509 phi->_inst_mem_id = new_idx;
510 no_of_updates++;
511 }
512 }
513
514 const Type* type = _new_type_array.fast_lookup(n->_idx);
515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
516 if (!_is_pass_finished) {
517 return -1; // delay
518 }
519 int old_idx = type->is_oopptr()->instance_id();
520 int new_idx = new_index(old_idx);
521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
522 _new_type_array.map(n->_idx, new_type);
523 no_of_updates++;
524 }
525
526 return no_of_updates;
527 }
528
529 void PhaseValues::init_con_caches() {
530 memset(_icons,0,sizeof(_icons));
531 memset(_lcons,0,sizeof(_lcons));
532 memset(_zcons,0,sizeof(_zcons));
533 }
534
535 //--------------------------------find_int_type--------------------------------
536 const TypeInt* PhaseValues::find_int_type(Node* n) {
537 if (n == nullptr) return nullptr;
538 // Call type_or_null(n) to determine node's type since we might be in
539 // parse phase and call n->Value() may return wrong type.
540 // (For example, a phi node at the beginning of loop parsing is not ready.)
541 const Type* t = type_or_null(n);
542 if (t == nullptr) return nullptr;
543 return t->isa_int();
544 }
545
546
547 //-------------------------------find_long_type--------------------------------
548 const TypeLong* PhaseValues::find_long_type(Node* n) {
549 if (n == nullptr) return nullptr;
550 // (See comment above on type_or_null.)
551 const Type* t = type_or_null(n);
552 if (t == nullptr) return nullptr;
553 return t->isa_long();
554 }
555
556 //------------------------------~PhaseValues-----------------------------------
557 #ifndef PRODUCT
558 PhaseValues::~PhaseValues() {
559 // Statistics for NodeHash
560 _table.dump();
561 // Statistics for value progress and efficiency
562 if( PrintCompilation && Verbose && WizardMode ) {
563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
565 if( made_transforms() != 0 ) {
566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
567 } else {
568 tty->cr();
569 }
570 }
571 }
572 #endif
573
574 //------------------------------makecon----------------------------------------
575 ConNode* PhaseValues::makecon(const Type* t) {
576 assert(t->singleton(), "must be a constant");
577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
578 switch (t->base()) { // fast paths
579 case Type::Half:
580 case Type::Top: return (ConNode*) C->top();
581 case Type::Int: return intcon( t->is_int()->get_con() );
582 case Type::Long: return longcon( t->is_long()->get_con() );
583 default: break;
584 }
585 if (t->is_zero_type())
586 return zerocon(t->basic_type());
587 return uncached_makecon(t);
588 }
589
590 //--------------------------uncached_makecon-----------------------------------
591 // Make an idealized constant - one of ConINode, ConPNode, etc.
592 ConNode* PhaseValues::uncached_makecon(const Type *t) {
593 assert(t->singleton(), "must be a constant");
594 ConNode* x = ConNode::make(t);
595 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
596 if (k == nullptr) {
597 set_type(x, t); // Missed, provide type mapping
598 GrowableArray<Node_Notes*>* nna = C->node_note_array();
599 if (nna != nullptr) {
600 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
601 loc->clear(); // do not put debug info on constants
602 }
603 } else {
604 x->destruct(this); // Hit, destroy duplicate constant
605 x = k; // use existing constant
606 }
607 return x;
608 }
609
610 //------------------------------intcon-----------------------------------------
611 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
612 ConINode* PhaseValues::intcon(jint i) {
613 // Small integer? Check cache! Check that cached node is not dead
614 if (i >= _icon_min && i <= _icon_max) {
615 ConINode* icon = _icons[i-_icon_min];
616 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
617 return icon;
618 }
619 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
620 assert(icon->is_Con(), "");
621 if (i >= _icon_min && i <= _icon_max)
622 _icons[i-_icon_min] = icon; // Cache small integers
623 return icon;
624 }
625
626 //------------------------------longcon----------------------------------------
627 // Fast long constant.
628 ConLNode* PhaseValues::longcon(jlong l) {
629 // Small integer? Check cache! Check that cached node is not dead
630 if (l >= _lcon_min && l <= _lcon_max) {
631 ConLNode* lcon = _lcons[l-_lcon_min];
632 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
633 return lcon;
634 }
635 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
636 assert(lcon->is_Con(), "");
637 if (l >= _lcon_min && l <= _lcon_max)
638 _lcons[l-_lcon_min] = lcon; // Cache small integers
639 return lcon;
640 }
641 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
642 if (bt == T_INT) {
643 return intcon(checked_cast<jint>(l));
644 }
645 assert(bt == T_LONG, "not an integer");
646 return longcon(l);
647 }
648
649
650 //------------------------------zerocon-----------------------------------------
651 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
652 ConNode* PhaseValues::zerocon(BasicType bt) {
653 assert((uint)bt <= _zcon_max, "domain check");
654 ConNode* zcon = _zcons[bt];
655 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
656 return zcon;
657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
658 _zcons[bt] = zcon;
659 return zcon;
660 }
661
662
663
664 //=============================================================================
665 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
666 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
667 if (i == nullptr) {
668 i = k->Ideal(this, can_reshape);
669 }
670 return i;
671 }
672
673 //------------------------------transform--------------------------------------
674 // Return a node which computes the same function as this node, but
675 // in a faster or cheaper fashion.
676 Node* PhaseGVN::transform(Node* n) {
677 NOT_PRODUCT( set_transforms(); )
678
679 // Apply the Ideal call in a loop until it no longer applies
680 Node* k = n;
681 Node* i = apply_ideal(k, /*can_reshape=*/false);
682 NOT_PRODUCT(uint loop_count = 1;)
683 while (i != nullptr) {
684 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
685 k = i;
686 #ifdef ASSERT
687 if (loop_count >= K + C->live_nodes()) {
688 dump_infinite_loop_info(i, "PhaseGVN::transform");
689 }
690 #endif
691 i = apply_ideal(k, /*can_reshape=*/false);
692 NOT_PRODUCT(loop_count++;)
693 }
694 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
695
696 // If brand new node, make space in type array.
697 ensure_type_or_null(k);
698
699 // Since I just called 'Value' to compute the set of run-time values
700 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
701 // cache Value. Later requests for the local phase->type of this Node can
702 // use the cached Value instead of suffering with 'bottom_type'.
703 const Type* t = k->Value(this); // Get runtime Value set
704 assert(t != nullptr, "value sanity");
705 if (type_or_null(k) != t) {
706 #ifndef PRODUCT
707 // Do not count initial visit to node as a transformation
708 if (type_or_null(k) == nullptr) {
709 inc_new_values();
710 set_progress();
711 }
712 #endif
713 set_type(k, t);
714 // If k is a TypeNode, capture any more-precise type permanently into Node
715 k->raise_bottom_type(t);
716 }
717
718 if (t->singleton() && !k->is_Con()) {
719 NOT_PRODUCT(set_progress();)
720 return makecon(t); // Turn into a constant
721 }
722
723 // Now check for Identities
724 i = k->Identity(this); // Look for a nearby replacement
725 if (i != k) { // Found? Return replacement!
726 NOT_PRODUCT(set_progress();)
727 return i;
728 }
729
730 // Global Value Numbering
731 i = hash_find_insert(k); // Insert if new
732 if (i && (i != k)) {
733 // Return the pre-existing node
734 NOT_PRODUCT(set_progress();)
735 return i;
736 }
737
738 // Return Idealized original
739 return k;
740 }
741
742 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
743 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
744 return false;
745 }
746 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
747 return false;
748 }
749 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
750 int i = 0;
751 while (d != n) {
752 n = IfNode::up_one_dom(n, linear_only);
753 i++;
754 if (n == nullptr || i >= 100) {
755 return false;
756 }
757 }
758 return true;
759 }
760
761 #ifdef ASSERT
762 //------------------------------dead_loop_check--------------------------------
763 // Check for a simple dead loop when a data node references itself directly
764 // or through an other data node excluding cons and phis.
765 void PhaseGVN::dead_loop_check( Node *n ) {
766 // Phi may reference itself in a loop
767 if (n != nullptr && !n->is_dead_loop_safe() && !n->is_CFG()) {
768 // Do 2 levels check and only data inputs.
769 bool no_dead_loop = true;
770 uint cnt = n->req();
771 for (uint i = 1; i < cnt && no_dead_loop; i++) {
772 Node *in = n->in(i);
773 if (in == n) {
774 no_dead_loop = false;
775 } else if (in != nullptr && !in->is_dead_loop_safe()) {
776 uint icnt = in->req();
777 for (uint j = 1; j < icnt && no_dead_loop; j++) {
778 if (in->in(j) == n || in->in(j) == in)
779 no_dead_loop = false;
780 }
781 }
782 }
783 if (!no_dead_loop) { n->dump_bfs(100, nullptr, ""); }
784 assert(no_dead_loop, "dead loop detected");
785 }
786 }
787
788
789 /**
790 * Dumps information that can help to debug the problem. A debug
791 * build fails with an assert.
792 */
793 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
794 n->dump(4);
795 assert(false, "infinite loop in %s", where);
796 }
797 #endif
798
799 //=============================================================================
800 //------------------------------PhaseIterGVN-----------------------------------
801 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
802 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
803 _worklist(*C->igvn_worklist())
804 {
805 _iterGVN = true;
806 assert(&_worklist == &igvn->_worklist, "sanity");
807 }
808
809 //------------------------------PhaseIterGVN-----------------------------------
810 // Initialize from scratch
811 PhaseIterGVN::PhaseIterGVN() : _delay_transform(false),
812 _worklist(*C->igvn_worklist())
813 {
814 _iterGVN = true;
815 uint max;
816
817 // Dead nodes in the hash table inherited from GVN were not treated as
818 // roots during def-use info creation; hence they represent an invisible
819 // use. Clear them out.
820 max = _table.size();
821 for( uint i = 0; i < max; ++i ) {
822 Node *n = _table.at(i);
823 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
824 if( n->is_top() ) continue;
825 // If remove_useless_nodes() has run, we expect no such nodes left.
826 assert(false, "remove_useless_nodes missed this node");
827 hash_delete(n);
828 }
829 }
830
831 // Any Phis or Regions on the worklist probably had uses that could not
832 // make more progress because the uses were made while the Phis and Regions
833 // were in half-built states. Put all uses of Phis and Regions on worklist.
834 max = _worklist.size();
835 for( uint j = 0; j < max; j++ ) {
836 Node *n = _worklist.at(j);
837 uint uop = n->Opcode();
838 if( uop == Op_Phi || uop == Op_Region ||
839 n->is_Type() ||
840 n->is_Mem() )
841 add_users_to_worklist(n);
842 }
843 }
844
845 void PhaseIterGVN::shuffle_worklist() {
846 if (_worklist.size() < 2) return;
847 for (uint i = _worklist.size() - 1; i >= 1; i--) {
848 uint j = C->random() % (i + 1);
849 swap(_worklist.adr()[i], _worklist.adr()[j]);
850 }
851 }
852
853 #ifndef PRODUCT
854 void PhaseIterGVN::verify_step(Node* n) {
855 if (is_verify_def_use()) {
856 ResourceMark rm;
857 VectorSet visited;
858 Node_List worklist;
859
860 _verify_window[_verify_counter % _verify_window_size] = n;
861 ++_verify_counter;
862 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
863 ++_verify_full_passes;
864 worklist.push(C->root());
865 Node::verify(-1, visited, worklist);
866 return;
867 }
868 for (int i = 0; i < _verify_window_size; i++) {
869 Node* n = _verify_window[i];
870 if (n == nullptr) {
871 continue;
872 }
873 if (n->in(0) == NodeSentinel) { // xform_idom
874 _verify_window[i] = n->in(1);
875 --i;
876 continue;
877 }
878 // Typical fanout is 1-2, so this call visits about 6 nodes.
879 if (!visited.test_set(n->_idx)) {
880 worklist.push(n);
881 }
882 }
883 Node::verify(4, visited, worklist);
884 }
885 }
886
887 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
888 const Type* newtype = type_or_null(n);
889 if (nn != n || oldtype != newtype) {
890 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
891 }
892 if (TraceIterativeGVN) {
893 uint wlsize = _worklist.size();
894 if (nn != n) {
895 // print old node
896 tty->print("< ");
897 if (oldtype != newtype && oldtype != nullptr) {
898 oldtype->dump();
899 }
900 do { tty->print("\t"); } while (tty->position() < 16);
901 tty->print("<");
902 n->dump();
903 }
904 if (oldtype != newtype || nn != n) {
905 // print new node and/or new type
906 if (oldtype == nullptr) {
907 tty->print("* ");
908 } else if (nn != n) {
909 tty->print("> ");
910 } else {
911 tty->print("= ");
912 }
913 if (newtype == nullptr) {
914 tty->print("null");
915 } else {
916 newtype->dump();
917 }
918 do { tty->print("\t"); } while (tty->position() < 16);
919 nn->dump();
920 }
921 if (Verbose && wlsize < _worklist.size()) {
922 tty->print(" Push {");
923 while (wlsize != _worklist.size()) {
924 Node* pushed = _worklist.at(wlsize++);
925 tty->print(" %d", pushed->_idx);
926 }
927 tty->print_cr(" }");
928 }
929 if (nn != n) {
930 // ignore n, it might be subsumed
931 verify_step((Node*) nullptr);
932 }
933 }
934 }
935
936 void PhaseIterGVN::init_verifyPhaseIterGVN() {
937 _verify_counter = 0;
938 _verify_full_passes = 0;
939 for (int i = 0; i < _verify_window_size; i++) {
940 _verify_window[i] = nullptr;
941 }
942 #ifdef ASSERT
943 // Verify that all modified nodes are on _worklist
944 Unique_Node_List* modified_list = C->modified_nodes();
945 while (modified_list != nullptr && modified_list->size()) {
946 Node* n = modified_list->pop();
947 if (!n->is_Con() && !_worklist.member(n)) {
948 n->dump();
949 fatal("modified node is not on IGVN._worklist");
950 }
951 }
952 #endif
953 }
954
955 void PhaseIterGVN::verify_PhaseIterGVN() {
956 #ifdef ASSERT
957 // Verify nodes with changed inputs.
958 Unique_Node_List* modified_list = C->modified_nodes();
959 while (modified_list != nullptr && modified_list->size()) {
960 Node* n = modified_list->pop();
961 if (!n->is_Con()) { // skip Con nodes
962 n->dump();
963 fatal("modified node was not processed by IGVN.transform_old()");
964 }
965 }
966 #endif
967
968 C->verify_graph_edges();
969 if (is_verify_def_use() && PrintOpto) {
970 if (_verify_counter == _verify_full_passes) {
971 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
972 (int) _verify_full_passes);
973 } else {
974 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
975 (int) _verify_counter, (int) _verify_full_passes);
976 }
977 }
978
979 #ifdef ASSERT
980 if (modified_list != nullptr) {
981 while (modified_list->size() > 0) {
982 Node* n = modified_list->pop();
983 n->dump();
984 assert(false, "VerifyIterativeGVN: new modified node was added");
985 }
986 }
987
988 verify_optimize();
989 #endif
990 }
991 #endif /* PRODUCT */
992
993 #ifdef ASSERT
994 /**
995 * Dumps information that can help to debug the problem. A debug
996 * build fails with an assert.
997 */
998 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
999 n->dump(4);
1000 _worklist.dump();
1001 assert(false, "infinite loop in %s", where);
1002 }
1003
1004 /**
1005 * Prints out information about IGVN if the 'verbose' option is used.
1006 */
1007 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1008 if (TraceIterativeGVN && Verbose) {
1009 tty->print(" Pop ");
1010 n->dump();
1011 if ((num_processed % 100) == 0) {
1012 _worklist.print_set();
1013 }
1014 }
1015 }
1016 #endif /* ASSERT */
1017
1018 void PhaseIterGVN::optimize() {
1019 DEBUG_ONLY(uint num_processed = 0;)
1020 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1021 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1022 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1023 if (StressIGVN) {
1024 shuffle_worklist();
1025 }
1026
1027 // The node count check in the loop below (check_node_count) assumes that we
1028 // increase the live node count with at most
1029 // max_live_nodes_increase_per_iteration in between checks. If this
1030 // assumption does not hold, there is a risk that we exceed the max node
1031 // limit in between checks and trigger an assert during node creation.
1032 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 5;
1033
1034 uint loop_count = 0;
1035 // Pull from worklist and transform the node. If the node has changed,
1036 // update edge info and put uses on worklist.
1037 while (_worklist.size() > 0) {
1038 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1039 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1040 return;
1041 }
1042 Node* n = _worklist.pop();
1043 if (loop_count >= K * C->live_nodes()) {
1044 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1045 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1046 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1047 return;
1048 }
1049 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1050 if (n->outcnt() != 0) {
1051 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1052 // Do the transformation
1053 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1054 Node* nn = transform_old(n);
1055 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1056 // Ensure we did not increase the live node count with more than
1057 // max_live_nodes_increase_per_iteration during the call to transform_old
1058 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1059 assert(increase < max_live_nodes_increase_per_iteration,
1060 "excessive live node increase in single iteration of IGVN: %d "
1061 "(should be at most %d)",
1062 increase, max_live_nodes_increase_per_iteration);
1063 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1064 } else if (!n->is_top()) {
1065 remove_dead_node(n);
1066 }
1067 loop_count++;
1068 }
1069 NOT_PRODUCT(verify_PhaseIterGVN();)
1070 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1071 }
1072
1073 #ifdef ASSERT
1074 void PhaseIterGVN::verify_optimize() {
1075 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1076
1077 if (is_verify_Value() ||
1078 is_verify_Ideal() ||
1079 is_verify_Identity() ||
1080 is_verify_invariants()) {
1081 ResourceMark rm;
1082 Unique_Node_List worklist;
1083 bool failure = false;
1084 // BFS all nodes, starting at root
1085 worklist.push(C->root());
1086 for (uint j = 0; j < worklist.size(); ++j) {
1087 Node* n = worklist.at(j);
1088 if (is_verify_Value()) { failure |= verify_Value_for(n); }
1089 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, false); }
1090 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, true); }
1091 if (is_verify_Identity()) { failure |= verify_Identity_for(n); }
1092 if (is_verify_invariants()) { failure |= verify_node_invariants_for(n); }
1093 // traverse all inputs and outputs
1094 for (uint i = 0; i < n->req(); i++) {
1095 if (n->in(i) != nullptr) {
1096 worklist.push(n->in(i));
1097 }
1098 }
1099 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1100 worklist.push(n->fast_out(i));
1101 }
1102 }
1103 // If we get this assert, check why the reported nodes were not processed again in IGVN.
1104 // We should either make sure that these nodes are properly added back to the IGVN worklist
1105 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception
1106 // in the verification code above if that is not possible for some reason (like Load nodes).
1107 assert(!failure, "Missed optimization opportunity/broken graph in PhaseIterGVN");
1108 }
1109
1110 verify_empty_worklist(nullptr);
1111 }
1112
1113 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1114 // Verify that the igvn worklist is empty. If no optimization happened, then
1115 // nothing needs to be on the worklist.
1116 if (_worklist.size() == 0) { return; }
1117
1118 stringStream ss; // Print as a block without tty lock.
1119 for (uint j = 0; j < _worklist.size(); j++) {
1120 Node* n = _worklist.at(j);
1121 ss.print("igvn.worklist[%d] ", j);
1122 n->dump("\n", false, &ss);
1123 }
1124 if (_worklist.size() != 0 && node != nullptr) {
1125 ss.print_cr("Previously optimized:");
1126 node->dump("\n", false, &ss);
1127 }
1128 tty->print_cr("%s", ss.as_string());
1129 assert(false, "igvn worklist must still be empty after verify");
1130 }
1131
1132 // Check that type(n) == n->Value(), return true if we have a failure.
1133 // We have a list of exceptions, see detailed comments in code.
1134 // (1) Integer "widen" changes, but the range is the same.
1135 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1136 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1137 bool PhaseIterGVN::verify_Value_for(Node* n, bool strict) {
1138 // If we assert inside type(n), because the type is still a null, then maybe
1139 // the node never went through gvn.transform, which would be a bug.
1140 const Type* told = type(n);
1141 const Type* tnew = n->Value(this);
1142 if (told == tnew) {
1143 return false;
1144 }
1145 // Exception (1)
1146 // Integer "widen" changes, but range is the same.
1147 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1148 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1149 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1150 if (t0->lo_as_long() == t1->lo_as_long() &&
1151 t0->hi_as_long() == t1->hi_as_long()) {
1152 return false; // ignore integer widen
1153 }
1154 }
1155 // Exception (2)
1156 // LoadNode performs deep traversals. Load is not notified for changes far away.
1157 if (!strict && n->is_Load() && !told->singleton()) {
1158 // MemNode::can_see_stored_value looks up through many memory nodes,
1159 // which means we would need to notify modifications from far up in
1160 // the inputs all the way down to the LoadNode. We don't do that.
1161 return false;
1162 }
1163 // Exception (3)
1164 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1165 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1166 // SubNode::Value
1167 // CmpPNode::sub
1168 // MemNode::detect_ptr_independence
1169 // MemNode::all_controls_dominate
1170 // We find all controls of a pointer load, and see if they dominate the control of
1171 // an allocation. If they all dominate, we know the allocation is after (independent)
1172 // of the pointer load, and we can say the pointers are different. For this we call
1173 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1174 // control sub of the allocation. The problems is that sometimes dominates answers
1175 // false conservatively, and later it can determine that it is indeed true. Loops with
1176 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1177 // so the traversal becomes more powerful. This is difficult to remidy, we would have
1178 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1179 // after loop-opts, so that should take care of many of these cases.
1180 return false;
1181 }
1182
1183 stringStream ss; // Print as a block without tty lock.
1184 ss.cr();
1185 ss.print_cr("Missed Value optimization:");
1186 n->dump_bfs(3, nullptr, "", &ss);
1187 ss.print_cr("Current type:");
1188 told->dump_on(&ss);
1189 ss.cr();
1190 ss.print_cr("Optimized type:");
1191 tnew->dump_on(&ss);
1192 ss.cr();
1193 tty->print_cr("%s", ss.as_string());
1194 return true;
1195 }
1196
1197 // Check that all Ideal optimizations that could be done were done.
1198 // Returns true if it found missed optimization opportunities and
1199 // false otherwise (no missed optimization, or skipped verification).
1200 bool PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1201 // First, we check a list of exceptions, where we skip verification,
1202 // because there are known cases where Ideal can optimize after IGVN.
1203 // Some may be expected and cannot be fixed, and others should be fixed.
1204 switch (n->Opcode()) {
1205 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1206 // (see "Range-Check scan limit"). So, it is possible that something
1207 // is optimized in that input subgraph, and the RangeCheck was not
1208 // added to the worklist because it would be too expensive to walk
1209 // down the graph for 1000 nodes and put all on the worklist.
1210 //
1211 // Found with:
1212 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1213 case Op_RangeCheck:
1214 return false;
1215
1216 // IfNode::Ideal does:
1217 // Node* prev_dom = search_identical(dist, igvn);
1218 // which means we seach up the CFG, traversing at most up to a distance.
1219 // If anything happens rather far away from the If, we may not put the If
1220 // back on the worklist.
1221 //
1222 // Found with:
1223 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1224 case Op_If:
1225 return false;
1226
1227 // IfNode::simple_subsuming
1228 // Looks for dominating test that subsumes the current test.
1229 // Notification could be difficult because of larger distance.
1230 //
1231 // Found with:
1232 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1233 // -XX:VerifyIterativeGVN=1110
1234 case Op_CountedLoopEnd:
1235 return false;
1236
1237 // LongCountedLoopEndNode::Ideal
1238 // Probably same issue as above.
1239 //
1240 // Found with:
1241 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1242 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1243 case Op_LongCountedLoopEnd:
1244 return false;
1245
1246 // RegionNode::Ideal does "Skip around the useless IF diamond".
1247 // 245 IfTrue === 244
1248 // 258 If === 245 257
1249 // 259 IfTrue === 258 [[ 263 ]]
1250 // 260 IfFalse === 258 [[ 263 ]]
1251 // 263 Region === 263 260 259 [[ 263 268 ]]
1252 // to
1253 // 245 IfTrue === 244
1254 // 263 Region === 263 245 _ [[ 263 268 ]]
1255 //
1256 // "Useless" means that there is no code in either branch of the If.
1257 // I found a case where this was not done yet during IGVN.
1258 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1259 //
1260 // Found with:
1261 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1262 case Op_Region:
1263 return false;
1264
1265 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1266 // that smaller idx are first. Tracking it back, it led me to
1267 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1268 //
1269 // Example:
1270 // Before PhaseIdealLoop::remix_address_expressions
1271 // 154 AddI === _ 12 144
1272 // After PhaseIdealLoop::remix_address_expressions
1273 // 154 AddI === _ 144 12
1274 // After AddNode::Ideal
1275 // 154 AddI === _ 12 144
1276 //
1277 // I suspect that the node should be added to the IGVN worklist after
1278 // PhaseIdealLoop::remix_address_expressions
1279 //
1280 // This is the only case I looked at, there may be others. Found like this:
1281 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1282 //
1283 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1284 //
1285 // Note: currently all of these fail also for other reasons, for example
1286 // because of "commute" doing the reordering with the phi below. Once
1287 // that is resolved, we can come back to this issue here.
1288 //
1289 // case Op_AddD:
1290 // case Op_AddI:
1291 // case Op_AddL:
1292 // case Op_AddF:
1293 // case Op_MulI:
1294 // case Op_MulL:
1295 // case Op_MulF:
1296 // case Op_MulD:
1297 // if (n->in(1)->_idx > n->in(2)->_idx) {
1298 // // Expect "commute" to revert this case.
1299 // return false;
1300 // }
1301 // break; // keep verifying
1302
1303 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1304 // Check for tight loop increments: Loop-phi of Add of loop-phi
1305 // It wants to take the phi into in(1):
1306 // 471 Phi === 435 38 390
1307 // 390 AddF === _ 471 391
1308 //
1309 // Other Associative operators are also affected equally.
1310 //
1311 // Investigate why this does not happen earlier during IGVN.
1312 //
1313 // Found with:
1314 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1315 // -XX:VerifyIterativeGVN=1110
1316 case Op_AddD:
1317 //case Op_AddI: // Also affected for other reasons, see case further down.
1318 //case Op_AddL: // Also affected for other reasons, see case further down.
1319 case Op_AddF:
1320 case Op_MulI:
1321 case Op_MulL:
1322 case Op_MulF:
1323 case Op_MulD:
1324 case Op_MinF:
1325 case Op_MinD:
1326 case Op_MaxF:
1327 case Op_MaxD:
1328 // XorINode::Ideal
1329 // Found with:
1330 // compiler/intrinsics/chacha/TestChaCha20.java
1331 // -XX:VerifyIterativeGVN=1110
1332 case Op_XorI:
1333 case Op_XorL:
1334 // It seems we may have similar issues with the HF cases.
1335 // Found with aarch64:
1336 // compiler/vectorization/TestFloat16VectorOperations.java
1337 // -XX:VerifyIterativeGVN=1110
1338 case Op_AddHF:
1339 case Op_MulHF:
1340 case Op_MaxHF:
1341 case Op_MinHF:
1342 return false;
1343
1344 // In MulNode::Ideal the edges can be swapped to help value numbering:
1345 //
1346 // // We are OK if right is a constant, or right is a load and
1347 // // left is a non-constant.
1348 // if( !(t2->singleton() ||
1349 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1350 // if( t1->singleton() || // Left input is a constant?
1351 // // Otherwise, sort inputs (commutativity) to help value numbering.
1352 // (in(1)->_idx > in(2)->_idx) ) {
1353 // swap_edges(1, 2);
1354 //
1355 // Why was this not done earlier during IGVN?
1356 //
1357 // Found with:
1358 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1359 // -XX:VerifyIterativeGVN=1110
1360 case Op_AndI:
1361 // Same for AndL.
1362 // Found with:
1363 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1364 // -XX:VerifyIterativeGVN=1110
1365 case Op_AndL:
1366 return false;
1367
1368 // SubLNode::Ideal does transform like:
1369 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1370 //
1371 // In IGVN before verification:
1372 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1373 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1374 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1375 // 8016 ConL === 0 [[ 8383 ]] #long:0
1376 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1377 //
1378 // And then in verification:
1379 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1380 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1381 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1382 // 8016 ConL === 0 [[ 8383 ]] #long:0
1383 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1384 //
1385 // So the form changed from:
1386 // c1 - (y + [8423 ConvI2L])
1387 // to
1388 // c1 - (y + -2)
1389 // but the SubL was not added to the IGVN worklist. Investigate why.
1390 // There could be other issues too.
1391 //
1392 // There seems to be a related AddL IGVN optimization that triggers
1393 // the same SubL optimization, so investigate that too.
1394 //
1395 // Found with:
1396 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1397 case Op_SubL:
1398 return false;
1399
1400 // SubINode::Ideal does
1401 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1402 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1403 //
1404 // Investigate why this does not yet happen during IGVN.
1405 //
1406 // Found with:
1407 // test/hotspot/jtreg/compiler/c2/IVTest.java
1408 // -XX:VerifyIterativeGVN=1110
1409 case Op_SubI:
1410 return false;
1411
1412 // AddNode::IdealIL does transform like:
1413 // Convert x + (con - y) into "(x - y) + con"
1414 //
1415 // In IGVN before verification:
1416 // 8382 ConvI2L
1417 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1418 // 8383 SubL === _ 8381 8382
1419 // 8168 ConvI2L
1420 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1421 //
1422 // And then in verification:
1423 // 8424 AddL
1424 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1425 // 8383 SubL === _ 8016 8424
1426 // 8168 ConvI2L
1427 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1428 //
1429 // So the form changed from:
1430 // x + (ConvI2L(0) - [8382 ConvI2L])
1431 // to
1432 // x + (0 - [8424 AddL])
1433 // but the AddL was not added to the IGVN worklist. Investigate why.
1434 // There could be other issues, too. For example with "commute", see above.
1435 //
1436 // Found with:
1437 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1438 case Op_AddL:
1439 return false;
1440
1441 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1442 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1443 // record_for_cleanup(cmp, phase);
1444 // This verification code in the Ideal code creates new nodes, and checks
1445 // if they fold in unexpected ways. This means some nodes are created and
1446 // added to the worklist, even if the SubTypeCheck is not optimized. This
1447 // goes agains the assumption of the verification here, which assumes that
1448 // if the node is not optimized, then no new nodes should be created, and
1449 // also no nodes should be added to the worklist.
1450 // I see two options:
1451 // 1) forbid what verify_helper does, because for each Ideal call it
1452 // uses memory and that is suboptimal. But it is not clear how that
1453 // verification can be done otherwise.
1454 // 2) Special case the verification here. Probably the new nodes that
1455 // were just created are dead, i.e. they are not connected down to
1456 // root. We could verify that, and remove those nodes from the graph
1457 // by setting all their inputs to nullptr. And of course we would
1458 // have to remove those nodes from the worklist.
1459 // Maybe there are other options too, I did not dig much deeper yet.
1460 //
1461 // Found with:
1462 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1463 case Op_SubTypeCheck:
1464 return false;
1465
1466 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1467 // to other nodes: Conv, Add, Sub, Mul, And ...
1468 //
1469 // 107 ConI === 0 [[ ... ]] #int:2
1470 // 84 LoadRange === _ 7 83
1471 // 50 ConI === 0 [[ ... ]] #int:0
1472 // 549 LoopLimit === _ 50 84 107
1473 //
1474 // I stepped backward, to see how the node was generated, and I found that it was
1475 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1476 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1477 // that, and it seems we want to skip lowering it until after loop-opts, but never
1478 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1479 // could be other issues too.
1480 //
1481 // Fond with:
1482 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1483 case Op_LoopLimit:
1484 return false;
1485
1486 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1487 // "This optimization tries to find two or more inputs of phi with the same constant
1488 // value. It then splits them into a separate Phi, and according Region."
1489 //
1490 // Example:
1491 // 130 DecodeN === _ 129
1492 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1493 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1494 //
1495 // turns into:
1496 //
1497 // 50 ConP === 0 [[ 99 91 18 ]] #null
1498 // 130 DecodeN === _ 129 [[ 18 ]]
1499 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1500 //
1501 // We would have to investigate why this optimization does not happen during IGVN.
1502 // There could also be other issues - I did not investigate further yet.
1503 //
1504 // Found with:
1505 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1506 case Op_Phi:
1507 return false;
1508
1509 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1510 // For examle "The allocated object does not escape".
1511 //
1512 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1513 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1514 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1515 // worklist when this change happens?
1516 //
1517 // Found with:
1518 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1519 case Op_MemBarStoreStore:
1520 return false;
1521
1522 // ConvI2LNode::Ideal converts
1523 // 648 AddI === _ 583 645 [[ 661 ]]
1524 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1525 // into
1526 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1527 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1528 // 773 AddL === _ 771 772 [[ ]]
1529 //
1530 // We have to investigate why this does not happen during IGVN in this case.
1531 // There could also be other issues - I did not investigate further yet.
1532 //
1533 // Found with:
1534 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1535 case Op_ConvI2L:
1536 return false;
1537
1538 // AddNode::IdealIL can do this transform (and similar other ones):
1539 // Convert "a*b+a*c into a*(b+c)
1540 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1541 // during IGVN? There was a mutation for one of the MulI, and only
1542 // after that the pattern was as needed for the optimization. The MulI
1543 // was added to the IGVN worklist, but not the AddI. This probably
1544 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1545 //
1546 // Found with:
1547 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1548 // -XX:VerifyIterativeGVN=1110
1549 case Op_AddI:
1550 return false;
1551
1552 // ArrayCopyNode::Ideal
1553 // calls ArrayCopyNode::prepare_array_copy
1554 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1555 // is not expected, and we create a range int:0..-1
1556 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1557 // note: the type is already empty!
1558 // calls PhaseIterGVN::transform
1559 // calls PhaseIterGVN::transform_old
1560 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1561 // calls Unique_Node_List::push -> pushes TOP to worklist
1562 //
1563 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1564 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1565 //
1566 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1567 // return nullptr, indicating nothing happened?
1568 // Is it smart to do transform in Compile::constrained_convI2L, and then
1569 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1570 // Should we just allow TOP to land on the worklist, as an exception?
1571 //
1572 // Found with:
1573 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1574 // -XX:VerifyIterativeGVN=1110
1575 case Op_ArrayCopy:
1576 return false;
1577
1578 // CastLLNode::Ideal
1579 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1580 //
1581 // Could be a notification issue, where updates inputs of CastLL do not notify
1582 // down through SubL to CastLL.
1583 //
1584 // Found With:
1585 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1586 // -XX:VerifyIterativeGVN=1110
1587 case Op_CastLL:
1588 return false;
1589
1590 // Similar case happens to CastII
1591 //
1592 // Found With:
1593 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1594 // -XX:VerifyIterativeGVN=1110
1595 case Op_CastII:
1596 return false;
1597
1598 // MaxLNode::Ideal
1599 // calls AddNode::Ideal
1600 // calls commute -> decides to swap edges
1601 //
1602 // Another notification issue, because we check inputs of inputs?
1603 // MaxL -> Phi -> Loop
1604 // MaxL -> Phi -> MaxL
1605 //
1606 // Found with:
1607 // compiler/c2/irTests/TestIfMinMax.java
1608 // -XX:VerifyIterativeGVN=1110
1609 case Op_MaxL:
1610 case Op_MinL:
1611 return false;
1612
1613 // OrINode::Ideal
1614 // calls AddNode::Ideal
1615 // calls commute -> left is Load, right not -> commute.
1616 //
1617 // Not sure why notification does not work here, seems like
1618 // the depth is only 1, so it should work. Needs investigation.
1619 //
1620 // Found with:
1621 // compiler/codegen/TestCharVect2.java#id0
1622 // -XX:VerifyIterativeGVN=1110
1623 case Op_OrI:
1624 case Op_OrL:
1625 return false;
1626
1627 // Bool -> constant folded to 1.
1628 // Issue with notification?
1629 //
1630 // Found with:
1631 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1632 // -XX:VerifyIterativeGVN=1110
1633 case Op_Bool:
1634 return false;
1635
1636 // LShiftLNode::Ideal
1637 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1638 // Probably a notification issue.
1639 //
1640 // Found with:
1641 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1642 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1643 case Op_LShiftL:
1644 return false;
1645
1646 // LShiftINode::Ideal
1647 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1648 // Could be issue with notification of inputs of inputs
1649 //
1650 // Side-note: should cases like these not be shared between
1651 // LShiftI and LShiftL?
1652 //
1653 // Found with:
1654 // compiler/escapeAnalysis/Test6689060.java
1655 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1656 case Op_LShiftI:
1657 return false;
1658
1659 // AddPNode::Ideal seems to do set_req without removing lock first.
1660 // Found with various vector tests tier1-tier3.
1661 case Op_AddP:
1662 return false;
1663
1664 // StrIndexOfNode::Ideal
1665 // Found in tier1-3.
1666 case Op_StrIndexOf:
1667 case Op_StrIndexOfChar:
1668 return false;
1669
1670 // StrEqualsNode::Identity
1671 //
1672 // Found (linux x64 only?) with:
1673 // serviceability/sa/ClhsdbThreadContext.java
1674 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1675 // Note: The -XX:LockingMode option is not available anymore.
1676 case Op_StrEquals:
1677 return false;
1678
1679 // AryEqNode::Ideal
1680 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1681 //
1682 // Found with:
1683 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1684 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1685 case Op_AryEq:
1686 return false;
1687
1688 // MergeMemNode::Ideal
1689 // Found in tier1-3. Did not investigate further yet.
1690 case Op_MergeMem:
1691 return false;
1692
1693 // URShiftINode::Ideal
1694 // Found in tier1-3. Did not investigate further yet.
1695 case Op_URShiftI:
1696 return false;
1697
1698 // CMoveINode::Ideal
1699 // Found in tier1-3. Did not investigate further yet.
1700 case Op_CMoveI:
1701 return false;
1702
1703 // CmpPNode::Ideal calls isa_const_java_mirror
1704 // and generates new constant nodes, even if no progress is made.
1705 // We can probably rewrite this so that only types are generated.
1706 // It seems that object types are not hashed, we could investigate
1707 // if that is an option as well.
1708 //
1709 // Found with:
1710 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1711 case Op_CmpP:
1712 return false;
1713
1714 // MinINode::Ideal
1715 // Did not investigate, but there are some patterns that might
1716 // need more notification.
1717 case Op_MinI:
1718 case Op_MaxI: // preemptively removed it as well.
1719 return false;
1720 }
1721
1722 if (n->is_Load()) {
1723 // LoadNode::Ideal uses tries to find an earlier memory state, and
1724 // checks can_see_stored_value for it.
1725 //
1726 // Investigate why this was not already done during IGVN.
1727 // A similar issue happens with Identity.
1728 //
1729 // There seem to be other cases where loads go up some steps, like
1730 // LoadNode::Ideal going up 10x steps to find dominating load.
1731 //
1732 // Found with:
1733 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1734 // -XX:VerifyIterativeGVN=1110
1735 return false;
1736 }
1737
1738 if (n->is_Store()) {
1739 // StoreNode::Ideal can do this:
1740 // // Capture an unaliased, unconditional, simple store into an initializer.
1741 // // Or, if it is independent of the allocation, hoist it above the allocation.
1742 // That replaces the Store with a MergeMem.
1743 //
1744 // We have to investigate why this does not happen during IGVN in this case.
1745 // There could also be other issues - I did not investigate further yet.
1746 //
1747 // Found with:
1748 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1749 return false;
1750 }
1751
1752 if (n->is_Vector()) {
1753 // VectorNode::Ideal swaps edges, but only for ops
1754 // that are deemed commutable. But swap_edges
1755 // requires the hash to be invariant when the edges
1756 // are swapped, which is not implemented for these
1757 // vector nodes. This seems not to create any trouble
1758 // usually, but we can also get graphs where in the
1759 // end the nodes are not all commuted, so there is
1760 // definitively an issue here.
1761 //
1762 // Probably we have two options: kill the hash, or
1763 // properly make the hash commutation friendly.
1764 //
1765 // Found with:
1766 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1767 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1768 return false;
1769 }
1770
1771 if (n->is_Region()) {
1772 // LoopNode::Ideal calls RegionNode::Ideal.
1773 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1774 // But I got an issue because RegionNode::optimize_trichotomy
1775 // then modifies another node, and pushes nodes to the worklist
1776 // Not sure if this is ok, modifying another node like that.
1777 // Maybe it is, then we need to look into what to do with
1778 // the nodes that are now on the worklist, maybe just clear
1779 // them out again. But maybe modifying other nodes like that
1780 // is also bad design. In the end, we return nullptr for
1781 // the current CountedLoop. But the extra nodes on the worklist
1782 // trip the asserts later on.
1783 //
1784 // Found with:
1785 // compiler/eliminateAutobox/TestShortBoxing.java
1786 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1787 return false;
1788 }
1789
1790 if (n->is_CallJava()) {
1791 // CallStaticJavaNode::Ideal
1792 // Led to a crash:
1793 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1794 //
1795 // Did not investigate yet, could be a bug.
1796 // Or maybe it does not expect to be called during verification.
1797 //
1798 // Found with:
1799 // test/jdk/jdk/incubator/vector/VectorRuns.java
1800 // -XX:VerifyIterativeGVN=1110
1801
1802 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1803 // and possibly their subclasses.
1804 // During late inlining it can call CallJavaNode::register_for_late_inline
1805 // That means we do more rounds of late inlining, but might fail.
1806 // Then we do IGVN again, and register the node again for late inlining.
1807 // This creates an endless cycle. Everytime we try late inlining, we
1808 // are also creating more nodes, especially SafePoint and MergeMem.
1809 // These nodes are immediately rejected when the inlining fails in the
1810 // do_late_inline_check, but they still grow the memory, until we hit
1811 // the MemLimit and crash.
1812 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1813 // called repeatedly, and eventually we terminate. I fear this is not
1814 // a great assumption to make. We should investigate more.
1815 //
1816 // Found with:
1817 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1818 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1819 return false;
1820 }
1821
1822 // The number of nodes shoud not increase.
1823 uint old_unique = C->unique();
1824 // The hash of a node should not change, this would indicate different inputs
1825 uint old_hash = n->hash();
1826 Node* i = n->Ideal(this, can_reshape);
1827 // If there was no new Idealization, we are probably happy.
1828 if (i == nullptr) {
1829 if (old_unique < C->unique()) {
1830 stringStream ss; // Print as a block without tty lock.
1831 ss.cr();
1832 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1833 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1834 n->dump_bfs(1, nullptr, "", &ss);
1835 tty->print_cr("%s", ss.as_string());
1836 return true;
1837 }
1838
1839 if (old_hash != n->hash()) {
1840 stringStream ss; // Print as a block without tty lock.
1841 ss.cr();
1842 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1843 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1844 n->dump_bfs(1, nullptr, "", &ss);
1845 tty->print_cr("%s", ss.as_string());
1846 return true;
1847 }
1848
1849 verify_empty_worklist(n);
1850
1851 // Everything is good.
1852 return false;
1853 }
1854
1855 // We just saw a new Idealization which was not done during IGVN.
1856 stringStream ss; // Print as a block without tty lock.
1857 ss.cr();
1858 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1859 if (i == n) {
1860 ss.print_cr("The node was reshaped by Ideal.");
1861 } else {
1862 ss.print_cr("The node was replaced by Ideal.");
1863 ss.print_cr("Old node:");
1864 n->dump_bfs(1, nullptr, "", &ss);
1865 }
1866 ss.print_cr("The result after Ideal:");
1867 i->dump_bfs(1, nullptr, "", &ss);
1868 tty->print_cr("%s", ss.as_string());
1869 return true;
1870 }
1871
1872 // Check that all Identity optimizations that could be done were done.
1873 // Returns true if it found missed optimization opportunities and
1874 // false otherwise (no missed optimization, or skipped verification).
1875 bool PhaseIterGVN::verify_Identity_for(Node* n) {
1876 // First, we check a list of exceptions, where we skip verification,
1877 // because there are known cases where Ideal can optimize after IGVN.
1878 // Some may be expected and cannot be fixed, and others should be fixed.
1879 switch (n->Opcode()) {
1880 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1881 // after loopopts:
1882 // // Transforming long counted loops requires a safepoint node. Do not
1883 // // eliminate a safepoint until loop opts are over.
1884 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1885 //
1886 // I think the check for major_progress does delay it until after loopopts
1887 // but it does not ensure that the node is on the IGVN worklist after
1888 // loopopts. I think we should try to instead check for
1889 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1890 //
1891 // Found with:
1892 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1893 case Op_SafePoint:
1894 return false;
1895
1896 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1897 // does not record any other memory splits.
1898 //
1899 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1900 // never got called during IGVN for this node, investigate why.
1901 //
1902 // Found with:
1903 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1904 case Op_MergeMem:
1905 return false;
1906
1907 // ConstraintCastNode::Identity finds casts that are the same, except that
1908 // the control is "higher up", i.e. dominates. The call goes via
1909 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1910 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1911 // away from the cast, but within 100 idom steps, the cast may not be
1912 // put on the IGVN worklist any more.
1913 //
1914 // Found with:
1915 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1916 case Op_CastPP:
1917 case Op_CastII:
1918 case Op_CastLL:
1919 return false;
1920
1921 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1922 // checks dominator, which may be changed, but too far up for notification
1923 // to work.
1924 //
1925 // Found with:
1926 // compiler/c2/irTests/TestSkeletonPredicates.java
1927 // -XX:VerifyIterativeGVN=1110
1928 case Op_CheckCastPP:
1929 return false;
1930
1931 // In SubNode::Identity, we do:
1932 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1933 // In the example, the AddI had an input replaced, the AddI is
1934 // added to the IGVN worklist, but the SubI is one link further
1935 // down and is not added. I checked add_users_of_use_to_worklist
1936 // where I would expect the SubI would be added, and I cannot
1937 // find the pattern, only this one:
1938 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1939 //
1940 // Fix this "notification" issue and check if there are any other
1941 // issues.
1942 //
1943 // Found with:
1944 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1945 case Op_SubI:
1946 case Op_SubL:
1947 return false;
1948
1949 // PhiNode::Identity checks for patterns like:
1950 // r = (x != con) ? x : con;
1951 // that can be constant folded to "x".
1952 //
1953 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
1954 // I suspect there was some earlier change to one of the inputs, but
1955 // not all relevant outputs were put on the IGVN worklist.
1956 //
1957 // Found with:
1958 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1959 // -XX:VerifyIterativeGVN=1110
1960 case Op_Phi:
1961 return false;
1962
1963 // ConvI2LNode::Identity does
1964 // convert I2L(L2I(x)) => x
1965 //
1966 // Investigate why this did not already happen during IGVN.
1967 //
1968 // Found with:
1969 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
1970 // -XX:VerifyIterativeGVN=1110
1971 case Op_ConvI2L:
1972 return false;
1973
1974 // MaxNode::find_identity_operation
1975 // Finds patterns like Max(A, Max(A, B)) -> Max(A, B)
1976 // This can be a 2-hop search, so maybe notification is not
1977 // good enough.
1978 //
1979 // Found with:
1980 // compiler/codegen/TestBooleanVect.java
1981 // -XX:VerifyIterativeGVN=1110
1982 case Op_MaxL:
1983 case Op_MinL:
1984 case Op_MaxI:
1985 case Op_MinI:
1986 case Op_MaxF:
1987 case Op_MinF:
1988 case Op_MaxHF:
1989 case Op_MinHF:
1990 case Op_MaxD:
1991 case Op_MinD:
1992 return false;
1993
1994
1995 // AddINode::Identity
1996 // Converts (x-y)+y to x
1997 // Could be issue with notification
1998 //
1999 // Turns out AddL does the same.
2000 //
2001 // Found with:
2002 // compiler/c2/Test6792161.java
2003 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2004 case Op_AddI:
2005 case Op_AddL:
2006 return false;
2007
2008 // AbsINode::Identity
2009 // Not investigated yet.
2010 case Op_AbsI:
2011 return false;
2012 }
2013
2014 if (n->is_Load()) {
2015 // LoadNode::Identity tries to look for an earlier store value via
2016 // can_see_stored_value. I found an example where this led to
2017 // an Allocation, where we could assume the value was still zero.
2018 // So the LoadN can be replaced with a zerocon.
2019 //
2020 // Investigate why this was not already done during IGVN.
2021 // A similar issue happens with Ideal.
2022 //
2023 // Found with:
2024 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2025 return false;
2026 }
2027
2028 if (n->is_Store()) {
2029 // StoreNode::Identity
2030 // Not investigated, but found missing optimization for StoreI.
2031 // Looks like a StoreI is replaced with an InitializeNode.
2032 //
2033 // Found with:
2034 // applications/ctw/modules/java_base_2.java
2035 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2036 return false;
2037 }
2038
2039 if (n->is_Vector()) {
2040 // Found with tier1-3. Not investigated yet.
2041 // The observed issue was with AndVNode::Identity
2042 return false;
2043 }
2044
2045 Node* i = n->Identity(this);
2046 // If we cannot find any other Identity, we are happy.
2047 if (i == n) {
2048 verify_empty_worklist(n);
2049 return false;
2050 }
2051
2052 // The verification just found a new Identity that was not found during IGVN.
2053 stringStream ss; // Print as a block without tty lock.
2054 ss.cr();
2055 ss.print_cr("Missed Identity optimization:");
2056 ss.print_cr("Old node:");
2057 n->dump_bfs(1, nullptr, "", &ss);
2058 ss.print_cr("New node:");
2059 i->dump_bfs(1, nullptr, "", &ss);
2060 tty->print_cr("%s", ss.as_string());
2061 return true;
2062 }
2063
2064 // Some other verifications that are not specific to a particular transformation.
2065 bool PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2066 if (n->is_AddP()) {
2067 if (!n->as_AddP()->address_input_has_same_base()) {
2068 stringStream ss; // Print as a block without tty lock.
2069 ss.cr();
2070 ss.print_cr("Base pointers must match for AddP chain:");
2071 n->dump_bfs(2, nullptr, "", &ss);
2072 tty->print_cr("%s", ss.as_string());
2073 return true;
2074 }
2075 }
2076 return false;
2077 }
2078 #endif
2079
2080 /**
2081 * Register a new node with the optimizer. Update the types array, the def-use
2082 * info. Put on worklist.
2083 */
2084 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2085 set_type_bottom(n);
2086 _worklist.push(n);
2087 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2088 return n;
2089 }
2090
2091 //------------------------------transform--------------------------------------
2092 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2093 Node *PhaseIterGVN::transform( Node *n ) {
2094 // If brand new node, make space in type array, and give it a type.
2095 ensure_type_or_null(n);
2096 if (type_or_null(n) == nullptr) {
2097 set_type_bottom(n);
2098 }
2099
2100 if (_delay_transform) {
2101 // Add the node to the worklist but don't optimize for now
2102 _worklist.push(n);
2103 return n;
2104 }
2105
2106 return transform_old(n);
2107 }
2108
2109 Node *PhaseIterGVN::transform_old(Node* n) {
2110 NOT_PRODUCT(set_transforms());
2111 // Remove 'n' from hash table in case it gets modified
2112 _table.hash_delete(n);
2113 #ifdef ASSERT
2114 if (is_verify_def_use()) {
2115 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2116 }
2117 #endif
2118
2119 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2120 if (n->is_Cmp()) {
2121 add_users_to_worklist(n);
2122 }
2123
2124 // Apply the Ideal call in a loop until it no longer applies
2125 Node* k = n;
2126 DEBUG_ONLY(dead_loop_check(k);)
2127 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2128 C->remove_modified_node(k);
2129 Node* i = apply_ideal(k, /*can_reshape=*/true);
2130 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2131 #ifndef PRODUCT
2132 verify_step(k);
2133 #endif
2134
2135 DEBUG_ONLY(uint loop_count = 1;)
2136 while (i != nullptr) {
2137 #ifdef ASSERT
2138 if (loop_count >= K + C->live_nodes()) {
2139 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2140 }
2141 #endif
2142 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2143 // Made a change; put users of original Node on worklist
2144 add_users_to_worklist(k);
2145 // Replacing root of transform tree?
2146 if (k != i) {
2147 // Make users of old Node now use new.
2148 subsume_node(k, i);
2149 k = i;
2150 }
2151 DEBUG_ONLY(dead_loop_check(k);)
2152 // Try idealizing again
2153 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2154 C->remove_modified_node(k);
2155 i = apply_ideal(k, /*can_reshape=*/true);
2156 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2157 #ifndef PRODUCT
2158 verify_step(k);
2159 #endif
2160 DEBUG_ONLY(loop_count++;)
2161 }
2162
2163 // If brand new node, make space in type array.
2164 ensure_type_or_null(k);
2165
2166 // See what kind of values 'k' takes on at runtime
2167 const Type* t = k->Value(this);
2168 assert(t != nullptr, "value sanity");
2169
2170 // Since I just called 'Value' to compute the set of run-time values
2171 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2172 // cache Value. Later requests for the local phase->type of this Node can
2173 // use the cached Value instead of suffering with 'bottom_type'.
2174 if (type_or_null(k) != t) {
2175 #ifndef PRODUCT
2176 inc_new_values();
2177 set_progress();
2178 #endif
2179 set_type(k, t);
2180 // If k is a TypeNode, capture any more-precise type permanently into Node
2181 k->raise_bottom_type(t);
2182 // Move users of node to worklist
2183 add_users_to_worklist(k);
2184 }
2185 // If 'k' computes a constant, replace it with a constant
2186 if (t->singleton() && !k->is_Con()) {
2187 NOT_PRODUCT(set_progress();)
2188 Node* con = makecon(t); // Make a constant
2189 add_users_to_worklist(k);
2190 subsume_node(k, con); // Everybody using k now uses con
2191 return con;
2192 }
2193
2194 // Now check for Identities
2195 i = k->Identity(this); // Look for a nearby replacement
2196 if (i != k) { // Found? Return replacement!
2197 NOT_PRODUCT(set_progress();)
2198 add_users_to_worklist(k);
2199 subsume_node(k, i); // Everybody using k now uses i
2200 return i;
2201 }
2202
2203 // Global Value Numbering
2204 i = hash_find_insert(k); // Check for pre-existing node
2205 if (i && (i != k)) {
2206 // Return the pre-existing node if it isn't dead
2207 NOT_PRODUCT(set_progress();)
2208 add_users_to_worklist(k);
2209 subsume_node(k, i); // Everybody using k now uses i
2210 return i;
2211 }
2212
2213 // Return Idealized original
2214 return k;
2215 }
2216
2217 //---------------------------------saturate------------------------------------
2218 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2219 const Type* limit_type) const {
2220 return new_type->narrow(old_type);
2221 }
2222
2223 //------------------------------remove_globally_dead_node----------------------
2224 // Kill a globally dead Node. All uses are also globally dead and are
2225 // aggressively trimmed.
2226 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2227 enum DeleteProgress {
2228 PROCESS_INPUTS,
2229 PROCESS_OUTPUTS
2230 };
2231 ResourceMark rm;
2232 Node_Stack stack(32);
2233 stack.push(dead, PROCESS_INPUTS);
2234
2235 while (stack.is_nonempty()) {
2236 dead = stack.node();
2237 if (dead->Opcode() == Op_SafePoint) {
2238 dead->as_SafePoint()->disconnect_from_root(this);
2239 }
2240 uint progress_state = stack.index();
2241 assert(dead != C->root(), "killing root, eh?");
2242 assert(!dead->is_top(), "add check for top when pushing");
2243 NOT_PRODUCT( set_progress(); )
2244 if (progress_state == PROCESS_INPUTS) {
2245 // After following inputs, continue to outputs
2246 stack.set_index(PROCESS_OUTPUTS);
2247 if (!dead->is_Con()) { // Don't kill cons but uses
2248 bool recurse = false;
2249 // Remove from hash table
2250 _table.hash_delete( dead );
2251 // Smash all inputs to 'dead', isolating him completely
2252 for (uint i = 0; i < dead->req(); i++) {
2253 Node *in = dead->in(i);
2254 if (in != nullptr && in != C->top()) { // Points to something?
2255 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2256 assert((nrep > 0), "sanity");
2257 if (in->outcnt() == 0) { // Made input go dead?
2258 stack.push(in, PROCESS_INPUTS); // Recursively remove
2259 recurse = true;
2260 } else if (in->outcnt() == 1 &&
2261 in->has_special_unique_user()) {
2262 _worklist.push(in->unique_out());
2263 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2264 if (in->Opcode() == Op_Region) {
2265 _worklist.push(in);
2266 } else if (in->is_Store()) {
2267 DUIterator_Fast imax, i = in->fast_outs(imax);
2268 _worklist.push(in->fast_out(i));
2269 i++;
2270 if (in->outcnt() == 2) {
2271 _worklist.push(in->fast_out(i));
2272 i++;
2273 }
2274 assert(!(i < imax), "sanity");
2275 }
2276 } else if (dead->is_data_proj_of_pure_function(in)) {
2277 _worklist.push(in);
2278 } else {
2279 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
2280 }
2281 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2282 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2283 // A Load that directly follows an InitializeNode is
2284 // going away. The Stores that follow are candidates
2285 // again to be captured by the InitializeNode.
2286 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
2287 Node *n = in->fast_out(j);
2288 if (n->is_Store()) {
2289 _worklist.push(n);
2290 }
2291 }
2292 }
2293 } // if (in != nullptr && in != C->top())
2294 } // for (uint i = 0; i < dead->req(); i++)
2295 if (recurse) {
2296 continue;
2297 }
2298 } // if (!dead->is_Con())
2299 } // if (progress_state == PROCESS_INPUTS)
2300
2301 // Aggressively kill globally dead uses
2302 // (Rather than pushing all the outs at once, we push one at a time,
2303 // plus the parent to resume later, because of the indefinite number
2304 // of edge deletions per loop trip.)
2305 if (dead->outcnt() > 0) {
2306 // Recursively remove output edges
2307 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2308 } else {
2309 // Finished disconnecting all input and output edges.
2310 stack.pop();
2311 // Remove dead node from iterative worklist
2312 _worklist.remove(dead);
2313 C->remove_useless_node(dead);
2314 }
2315 } // while (stack.is_nonempty())
2316 }
2317
2318 //------------------------------subsume_node-----------------------------------
2319 // Remove users from node 'old' and add them to node 'nn'.
2320 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2321 if (old->Opcode() == Op_SafePoint) {
2322 old->as_SafePoint()->disconnect_from_root(this);
2323 }
2324 assert( old != hash_find(old), "should already been removed" );
2325 assert( old != C->top(), "cannot subsume top node");
2326 // Copy debug or profile information to the new version:
2327 C->copy_node_notes_to(nn, old);
2328 // Move users of node 'old' to node 'nn'
2329 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2330 Node* use = old->last_out(i); // for each use...
2331 // use might need re-hashing (but it won't if it's a new node)
2332 rehash_node_delayed(use);
2333 // Update use-def info as well
2334 // We remove all occurrences of old within use->in,
2335 // so as to avoid rehashing any node more than once.
2336 // The hash table probe swamps any outer loop overhead.
2337 uint num_edges = 0;
2338 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2339 if (use->in(j) == old) {
2340 use->set_req(j, nn);
2341 ++num_edges;
2342 }
2343 }
2344 i -= num_edges; // we deleted 1 or more copies of this edge
2345 }
2346
2347 // Search for instance field data PhiNodes in the same region pointing to the old
2348 // memory PhiNode and update their instance memory ids to point to the new node.
2349 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2350 Node* region = old->in(0);
2351 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2352 PhiNode* phi = region->fast_out(i)->isa_Phi();
2353 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2354 phi->set_inst_mem_id((int)nn->_idx);
2355 }
2356 }
2357 }
2358
2359 // Smash all inputs to 'old', isolating him completely
2360 Node *temp = new Node(1);
2361 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2362 remove_dead_node( old );
2363 temp->del_req(0); // Yank bogus edge
2364 if (nn != nullptr && nn->outcnt() == 0) {
2365 _worklist.push(nn);
2366 }
2367 #ifndef PRODUCT
2368 if (is_verify_def_use()) {
2369 for ( int i = 0; i < _verify_window_size; i++ ) {
2370 if ( _verify_window[i] == old )
2371 _verify_window[i] = nn;
2372 }
2373 }
2374 #endif
2375 temp->destruct(this); // reuse the _idx of this little guy
2376 }
2377
2378 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) {
2379 assert(n != nullptr, "sanity");
2380 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2381 Node* u = n->fast_out(i);
2382 if (u != n) {
2383 rehash_node_delayed(u);
2384 int nb = u->replace_edge(n, m);
2385 --i, imax -= nb;
2386 }
2387 }
2388 assert(n->outcnt() == 0, "all uses must be deleted");
2389 }
2390
2391 //------------------------------add_users_to_worklist--------------------------
2392 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2393 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2394 worklist.push(n->fast_out(i)); // Push on worklist
2395 }
2396 }
2397
2398 // Return counted loop Phi if as a counted loop exit condition, cmp
2399 // compares the induction variable with n
2400 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2401 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2402 Node* bol = cmp->fast_out(i);
2403 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2404 Node* iff = bol->fast_out(i2);
2405 if (iff->is_BaseCountedLoopEnd()) {
2406 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2407 if (cle->limit() == n) {
2408 PhiNode* phi = cle->phi();
2409 if (phi != nullptr) {
2410 return phi;
2411 }
2412 }
2413 }
2414 }
2415 }
2416 return nullptr;
2417 }
2418
2419 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2420 add_users_to_worklist0(n, _worklist);
2421
2422 Unique_Node_List& worklist = _worklist;
2423 // Move users of node to worklist
2424 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2425 Node* use = n->fast_out(i); // Get use
2426 add_users_of_use_to_worklist(n, use, worklist);
2427 }
2428 }
2429
2430 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2431 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2432 use->is_Store() ) // Enable store/load same address
2433 add_users_to_worklist0(use, worklist);
2434
2435 // If we changed the receiver type to a call, we need to revisit
2436 // the Catch following the call. It's looking for a non-null
2437 // receiver to know when to enable the regular fall-through path
2438 // in addition to the NullPtrException path.
2439 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2440 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2441 if (p != nullptr) {
2442 add_users_to_worklist0(p, worklist);
2443 }
2444 }
2445
2446 // AndLNode::Ideal folds GraphKit::mark_word_test patterns. Give it a chance to run.
2447 if (n->is_Load() && use->is_Phi()) {
2448 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2449 Node* u = use->fast_out(i);
2450 if (u->Opcode() == Op_AndL) {
2451 worklist.push(u);
2452 }
2453 }
2454 }
2455
2456 uint use_op = use->Opcode();
2457 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2458 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2459 if (use->outcnt() > 0) {
2460 Node* bol = use->raw_out(0);
2461 if (bol->outcnt() > 0) {
2462 Node* iff = bol->raw_out(0);
2463 if (iff->outcnt() == 2) {
2464 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2465 // phi merging either 0 or 1 onto the worklist
2466 Node* ifproj0 = iff->raw_out(0);
2467 Node* ifproj1 = iff->raw_out(1);
2468 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2469 Node* region0 = ifproj0->raw_out(0);
2470 Node* region1 = ifproj1->raw_out(0);
2471 if( region0 == region1 )
2472 add_users_to_worklist0(region0, worklist);
2473 }
2474 }
2475 }
2476 }
2477 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2478 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2479 if (phi != nullptr) {
2480 // Input to the cmp of a loop exit check has changed, thus
2481 // the loop limit may have changed, which can then change the
2482 // range values of the trip-count Phi.
2483 worklist.push(phi);
2484 }
2485 }
2486 if (use_op == Op_CmpI) {
2487 Node* cmp = use;
2488 Node* in1 = cmp->in(1);
2489 Node* in2 = cmp->in(2);
2490 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2491 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2492 //
2493 // in1 in2 in1 in2
2494 // | | | |
2495 // +--- | --+ | |
2496 // | | | | |
2497 // CmpINode | CmpINode
2498 // | | |
2499 // BoolNode | BoolNode
2500 // | | OR |
2501 // IfNode | IfNode
2502 // | | |
2503 // IfProj | IfProj X
2504 // | | | |
2505 // CastIINode CastIINode
2506 //
2507 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2508 if (in1 == n) {
2509 // in1 modified -> could turn into X -> do traversal based on right pattern.
2510 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2511 Node* bol = cmp->fast_out(i2); // For each Bool
2512 if (bol->is_Bool()) {
2513 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2514 Node* iff = bol->fast_out(i3); // For each If
2515 if (iff->is_If()) {
2516 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2517 Node* if_proj = iff->fast_out(i4); // For each IfProj
2518 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2519 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2520 Node* castii = if_proj->fast_out(i5); // For each CastII
2521 if (castii->is_CastII() &&
2522 castii->as_CastII()->carry_dependency()) {
2523 worklist.push(castii);
2524 }
2525 }
2526 }
2527 }
2528 }
2529 }
2530 }
2531 } else {
2532 // Only in2 modified -> can assume X == in2 (left pattern).
2533 assert(n == in2, "only in2 modified");
2534 // Find all CastII with input in1.
2535 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2536 Node* castii = in1->fast_out(j);
2537 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2538 // Find If.
2539 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2540 Node* ifnode = castii->in(0)->in(0);
2541 // Check that if connects to the cmp
2542 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2543 worklist.push(castii);
2544 }
2545 }
2546 }
2547 }
2548 }
2549 }
2550 }
2551 }
2552
2553 // Inline type nodes can have other inline types as users. If an input gets
2554 // updated, make sure that inline type users get a chance for optimization.
2555 if (use->is_InlineType()) {
2556 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2557 Node* u = use->fast_out(i2);
2558 if (u->is_InlineType())
2559 worklist.push(u);
2560 }
2561 }
2562 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2563 // Patterns:
2564 // ConstraintCast+ -> Sub
2565 // ConstraintCast+ -> Phi
2566 // ConstraintCast+ -> Xor
2567 if (use->is_ConstraintCast()) {
2568 auto push_the_uses_to_worklist = [&](Node* n){
2569 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2570 worklist.push(n);
2571 }
2572 };
2573 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2574 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2575 }
2576 // If changed LShift inputs, check RShift users for useless sign-ext
2577 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2578 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2579 Node* u = use->fast_out(i2);
2580 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL)
2581 worklist.push(u);
2582 }
2583 }
2584 // If changed LShift inputs, check And users for shift and mask (And) operation
2585 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2586 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2587 Node* u = use->fast_out(i2);
2588 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
2589 worklist.push(u);
2590 }
2591 }
2592 }
2593 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2594 if (use_op == Op_AddI || use_op == Op_SubI) {
2595 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2596 Node* u = use->fast_out(i2);
2597 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
2598 worklist.push(u);
2599 }
2600 }
2601 }
2602 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2603 if (use_op == Op_AndI || use_op == Op_AndL) {
2604 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2605 Node* u = use->fast_out(i2);
2606 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2607 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL) {
2608 worklist.push(u);
2609 }
2610 }
2611 }
2612 // Check for redundant conversion patterns:
2613 // ConvD2L->ConvL2D->ConvD2L
2614 // ConvF2I->ConvI2F->ConvF2I
2615 // ConvF2L->ConvL2F->ConvF2L
2616 // ConvI2F->ConvF2I->ConvI2F
2617 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2618 // are the only ones that are known to trigger missed optimizations otherwise
2619 if (use_op == Op_ConvL2D ||
2620 use_op == Op_ConvI2F ||
2621 use_op == Op_ConvL2F ||
2622 use_op == Op_ConvF2I) {
2623 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2624 Node* u = use->fast_out(i2);
2625 if ((use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2626 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2627 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2628 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F)) {
2629 worklist.push(u);
2630 }
2631 }
2632 }
2633 // If changed AddP inputs:
2634 // - check Stores for loop invariant, and
2635 // - if the changed input is the offset, check constant-offset AddP users for
2636 // address expression flattening.
2637 if (use_op == Op_AddP) {
2638 bool offset_changed = n == use->in(AddPNode::Offset);
2639 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2640 Node* u = use->fast_out(i2);
2641 if (u->is_Mem()) {
2642 worklist.push(u);
2643 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) {
2644 worklist.push(u);
2645 }
2646 }
2647 }
2648 // Check for "abs(0-x)" into "abs(x)" conversion
2649 if (use->is_Sub()) {
2650 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2651 Node* u = use->fast_out(i2);
2652 if (u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2653 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI) {
2654 worklist.push(u);
2655 }
2656 }
2657 }
2658 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2659 add_users_to_worklist0(proj, worklist);
2660 };
2661 // If changed initialization activity, check dependent Stores
2662 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2663 InitializeNode* init = use->as_Allocate()->initialization();
2664 if (init != nullptr) {
2665 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2666 }
2667 }
2668 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2669 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2670 // to guarantee the change is not missed.
2671 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2672 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2673 if (p != nullptr) {
2674 add_users_to_worklist0(p, worklist);
2675 }
2676 }
2677
2678 if (use_op == Op_Initialize) {
2679 InitializeNode* init = use->as_Initialize();
2680 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2681 }
2682 // Loading the java mirror from a Klass requires two loads and the type
2683 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2684 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
2685 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2686 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
2687
2688 if (use_op == Op_CastP2X) {
2689 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2690 Node* u = use->fast_out(i2);
2691 // TODO 8350865 Still needed? Yes, I think this is from PhaseMacroExpand::expand_mh_intrinsic_return
2692 if (u->Opcode() == Op_AndX) {
2693 worklist.push(u);
2694 }
2695 // Search for CmpL(OrL(CastP2X(..), CastP2X(..)), 0L)
2696 if (u->Opcode() == Op_OrL) {
2697 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
2698 Node* cmp = u->fast_out(i3);
2699 if (cmp->Opcode() == Op_CmpL) {
2700 worklist.push(cmp);
2701 }
2702 }
2703 }
2704 }
2705 }
2706 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2707 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2708 Node* u = use->fast_out(i2);
2709 const Type* ut = u->bottom_type();
2710 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2711 if (has_load_barrier_nodes) {
2712 // Search for load barriers behind the load
2713 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
2714 Node* b = u->fast_out(i3);
2715 if (bs->is_gc_barrier_node(b)) {
2716 worklist.push(b);
2717 }
2718 }
2719 }
2720 worklist.push(u);
2721 }
2722 }
2723 }
2724 // Give CallStaticJavaNode::remove_useless_allocation a chance to run
2725 if (use->is_Region()) {
2726 Node* c = use;
2727 do {
2728 c = c->unique_ctrl_out_or_null();
2729 } while (c != nullptr && c->is_Region());
2730 if (c != nullptr && c->is_CallStaticJava() && c->as_CallStaticJava()->uncommon_trap_request() != 0) {
2731 worklist.push(c);
2732 }
2733 }
2734 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2735 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2736 if (use->outcnt() == 1) {
2737 Node* cmp = use->unique_out();
2738 worklist.push(cmp);
2739 }
2740 }
2741
2742 // From CastX2PNode::Ideal
2743 // CastX2P(AddX(x, y))
2744 // CastX2P(SubX(x, y))
2745 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2746 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2747 Node* u = use->fast_out(i2);
2748 if (u->Opcode() == Op_CastX2P) {
2749 worklist.push(u);
2750 }
2751 }
2752 }
2753
2754 /* AndNode has a special handling when one of the operands is a LShiftNode:
2755 * (LHS << s) & RHS
2756 * if RHS fits in less than s bits, the value of this expression is 0.
2757 * The difficulty is that there might be a conversion node (ConvI2L) between
2758 * the LShiftINode and the AndLNode, like so:
2759 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2760 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2761 * (see `AndIL_min_trailing_zeros`).
2762 *
2763 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2764 * is not enough: there might be no update happening there. We need to
2765 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2766 *
2767 * Moreover we can have ConstraintCasts in between. It may look like
2768 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2769 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2770 * So we must deal with that as well.
2771 */
2772 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2773 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2774 auto push_and_to_worklist = [&worklist](Node* n){
2775 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2776 worklist.push(n);
2777 }
2778 };
2779 use->visit_uses(push_and_to_worklist, is_boundary);
2780 }
2781 }
2782
2783 /**
2784 * Remove the speculative part of all types that we know of
2785 */
2786 void PhaseIterGVN::remove_speculative_types() {
2787 assert(UseTypeSpeculation, "speculation is off");
2788 for (uint i = 0; i < _types.Size(); i++) {
2789 const Type* t = _types.fast_lookup(i);
2790 if (t != nullptr) {
2791 _types.map(i, t->remove_speculative());
2792 }
2793 }
2794 _table.check_no_speculative_types();
2795 }
2796
2797 // Check if the type of a divisor of a Div or Mod node includes zero.
2798 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
2799 switch (n->Opcode()) {
2800 case Op_DivI:
2801 case Op_ModI:
2802 case Op_UDivI:
2803 case Op_UModI: {
2804 // Type of divisor includes 0?
2805 if (type(n->in(2)) == Type::TOP) {
2806 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2807 return false;
2808 }
2809 const TypeInt* type_divisor = type(n->in(2))->is_int();
2810 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2811 }
2812 case Op_DivL:
2813 case Op_ModL:
2814 case Op_UDivL:
2815 case Op_UModL: {
2816 // Type of divisor includes 0?
2817 if (type(n->in(2)) == Type::TOP) {
2818 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2819 return false;
2820 }
2821 const TypeLong* type_divisor = type(n->in(2))->is_long();
2822 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2823 }
2824 }
2825 return true;
2826 }
2827
2828 //=============================================================================
2829 #ifndef PRODUCT
2830 uint PhaseCCP::_total_invokes = 0;
2831 uint PhaseCCP::_total_constants = 0;
2832 #endif
2833 //------------------------------PhaseCCP---------------------------------------
2834 // Conditional Constant Propagation, ala Wegman & Zadeck
2835 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2836 NOT_PRODUCT( clear_constants(); )
2837 assert( _worklist.size() == 0, "" );
2838 analyze();
2839 }
2840
2841 #ifndef PRODUCT
2842 //------------------------------~PhaseCCP--------------------------------------
2843 PhaseCCP::~PhaseCCP() {
2844 inc_invokes();
2845 _total_constants += count_constants();
2846 }
2847 #endif
2848
2849
2850 #ifdef ASSERT
2851 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2852 if (tnew->meet(told) != tnew->remove_speculative()) {
2853 n->dump(3);
2854 tty->print("told = "); told->dump(); tty->cr();
2855 tty->print("tnew = "); tnew->dump(); tty->cr();
2856 fatal("Not monotonic");
2857 }
2858 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2859 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2860 }
2861 #endif //ASSERT
2862
2863 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2864 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2865 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2866 // or nodes even further down back to the worklist because their type could change as a result of the current type
2867 // change).
2868 void PhaseCCP::analyze() {
2869 // Initialize all types to TOP, optimistic analysis
2870 for (uint i = 0; i < C->unique(); i++) {
2871 _types.map(i, Type::TOP);
2872 }
2873
2874 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2875 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2876 // Compile is over. The local arena gets de-allocated at the end of its scope.
2877 ResourceArea local_arena(mtCompiler);
2878 Unique_Node_List worklist(&local_arena);
2879 Unique_Node_List worklist_revisit(&local_arena);
2880 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2881
2882 // Push root onto worklist
2883 worklist.push(C->root());
2884
2885 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2886 _root_and_safepoints.push(C->root());
2887
2888 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2889
2890 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2891 while (worklist.size() != 0) {
2892 Node* n = fetch_next_node(worklist);
2893 DEBUG_ONLY(worklist_verify.push(n);)
2894 if (needs_revisit(n)) {
2895 worklist_revisit.push(n);
2896 }
2897 if (n->is_SafePoint()) {
2898 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2899 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2900 _root_and_safepoints.push(n);
2901 }
2902 analyze_step(worklist, n);
2903 }
2904
2905 // More rounds to catch updates far in the graph.
2906 // Revisit nodes that might be able to refine their types at the end of the round.
2907 // If so, process these nodes. If there is remaining work, start another round.
2908 do {
2909 while (worklist.size() != 0) {
2910 Node* n = fetch_next_node(worklist);
2911 analyze_step(worklist, n);
2912 }
2913 for (uint t = 0; t < worklist_revisit.size(); t++) {
2914 Node* n = worklist_revisit.at(t);
2915 analyze_step(worklist, n);
2916 }
2917 } while (worklist.size() != 0);
2918
2919 DEBUG_ONLY(verify_analyze(worklist_verify);)
2920 }
2921
2922 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
2923 const Type* new_type = n->Value(this);
2924 if (new_type != type(n)) {
2925 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2926 dump_type_and_node(n, new_type);
2927 set_type(n, new_type);
2928 push_child_nodes_to_worklist(worklist, n);
2929 }
2930 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2931 // Keep track of Type nodes to kill CFG paths that use Type
2932 // nodes that become dead.
2933 _maybe_top_type_nodes.push(n);
2934 }
2935 }
2936
2937 // Some nodes can refine their types due to type change somewhere deep
2938 // in the graph. We will need to revisit them before claiming convergence.
2939 // Add nodes here if particular *Node::Value is doing deep graph traversals
2940 // not handled by PhaseCCP::push_more_uses().
2941 bool PhaseCCP::needs_revisit(Node* n) const {
2942 // LoadNode performs deep traversals. Load is not notified for changes far away.
2943 if (n->is_Load()) {
2944 return true;
2945 }
2946 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
2947 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
2948 return true;
2949 }
2950 return false;
2951 }
2952
2953 #ifdef ASSERT
2954 // For every node n on verify list, check if type(n) == n->Value()
2955 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
2956 // Therefore, we are verifying Value convergence strictly.
2957 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2958 bool failure = false;
2959 while (worklist_verify.size()) {
2960 Node* n = worklist_verify.pop();
2961 failure |= verify_Value_for(n, /* strict = */ true);
2962 }
2963 // If we get this assert, check why the reported nodes were not processed again in CCP.
2964 // We should either make sure that these nodes are properly added back to the CCP worklist
2965 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
2966 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
2967 // them at the end of the round.
2968 assert(!failure, "PhaseCCP not at fixpoint: analysis result may be unsound.");
2969 }
2970 #endif
2971
2972 // Fetch next node from worklist to be examined in this iteration.
2973 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2974 if (StressCCP) {
2975 return worklist.remove(C->random() % worklist.size());
2976 } else {
2977 return worklist.pop();
2978 }
2979 }
2980
2981 #ifndef PRODUCT
2982 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2983 if (TracePhaseCCP) {
2984 t->dump();
2985 do {
2986 tty->print("\t");
2987 } while (tty->position() < 16);
2988 n->dump();
2989 }
2990 }
2991 #endif
2992
2993 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2994 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2995 // of the new type of 'n'. Push these nodes to the worklist.
2996 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2997 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2998 Node* use = n->fast_out(i);
2999 push_if_not_bottom_type(worklist, use);
3000 push_more_uses(worklist, n, use);
3001 }
3002 }
3003
3004 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
3005 if (n->bottom_type() != type(n)) {
3006 worklist.push(n);
3007 }
3008 }
3009
3010 // For some nodes, we need to propagate the type change to grandchildren or even further down.
3011 // Add them back to the worklist.
3012 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
3013 push_phis(worklist, use);
3014 push_catch(worklist, use);
3015 push_cmpu(worklist, use);
3016 push_counted_loop_phi(worklist, parent, use);
3017 push_cast(worklist, use);
3018 push_loadp(worklist, use);
3019 push_and(worklist, parent, use);
3020 push_cast_ii(worklist, parent, use);
3021 push_opaque_zero_trip_guard(worklist, use);
3022 push_bool_with_cmpu_and_mask(worklist, use);
3023 }
3024
3025
3026 // We must recheck Phis too if use is a Region.
3027 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
3028 if (use->is_Region()) {
3029 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3030 push_if_not_bottom_type(worklist, use->fast_out(i));
3031 }
3032 }
3033 }
3034
3035 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
3036 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
3037 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
3038 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
3039 if (use->is_Call()) {
3040 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3041 Node* proj = use->fast_out(i);
3042 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
3043 Node* catch_node = proj->find_out_with(Op_Catch);
3044 if (catch_node != nullptr) {
3045 worklist.push(catch_node);
3046 }
3047 }
3048 }
3049 }
3050 }
3051
3052 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
3053 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
3054 // and get wrong types otherwise.
3055 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
3056 uint use_op = use->Opcode();
3057 if (use_op == Op_AddI || use_op == Op_SubI) {
3058 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3059 Node* cmpu = use->fast_out(i);
3060 const uint cmpu_opcode = cmpu->Opcode();
3061 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) {
3062 // Got a CmpU or CmpU3 which might need the new type information from node n.
3063 push_if_not_bottom_type(worklist, cmpu);
3064 }
3065 }
3066 }
3067 }
3068
3069 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
3070 // (1b): "(m & x) <u (m + 1))".
3071 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
3072 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
3073 //
3074 // m x m 1 (%%)
3075 // \ / \ /
3076 // AndI AddI
3077 // \ /
3078 // CmpU
3079 // |
3080 // Bool
3081 //
3082 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3083 uint use_op = use->Opcode();
3084 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3085 // Not "m & x" or "m + 1"
3086 return;
3087 }
3088 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3089 Node* cmpu = use->fast_out(i);
3090 if (cmpu->Opcode() == Op_CmpU) {
3091 push_bool_matching_case1b(worklist, cmpu);
3092 }
3093 }
3094 }
3095
3096 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3097 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3098 assert(cmpu->Opcode() == Op_CmpU, "must be");
3099 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3100 Node* bol = cmpu->fast_out(i);
3101 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3102 // Not a Bool with "<u"
3103 continue;
3104 }
3105 Node* andI = cmpu->in(1);
3106 Node* addI = cmpu->in(2);
3107 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3108 // Not "m & x" and "m + 1"
3109 continue;
3110 }
3111
3112 Node* m = addI->in(1);
3113 if (m == andI->in(1) || m == andI->in(2)) {
3114 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3115 push_if_not_bottom_type(worklist, bol);
3116 }
3117 }
3118 }
3119
3120 // If n is used in a counted loop exit condition, then the type of the counted loop's Phi depends on the type of 'n'.
3121 // Seem PhiNode::Value().
3122 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3123 uint use_op = use->Opcode();
3124 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3125 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3126 if (phi != nullptr) {
3127 worklist.push(phi);
3128 }
3129 }
3130 }
3131
3132 // TODO 8350865 Still needed? Yes, I think this is from PhaseMacroExpand::expand_mh_intrinsic_return
3133 void PhaseCCP::push_cast(Unique_Node_List& worklist, const Node* use) {
3134 uint use_op = use->Opcode();
3135 if (use_op == Op_CastP2X) {
3136 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
3137 Node* u = use->fast_out(i2);
3138 if (u->Opcode() == Op_AndX) {
3139 worklist.push(u);
3140 }
3141 }
3142 }
3143 }
3144
3145 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3146 // See LoadNode::Value().
3147 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3148 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
3149 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
3150
3151 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3152 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3153 Node* loadp = use->fast_out(i);
3154 const Type* ut = loadp->bottom_type();
3155 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3156 if (has_load_barrier_nodes) {
3157 // Search for load barriers behind the load
3158 push_load_barrier(worklist, barrier_set, loadp);
3159 }
3160 worklist.push(loadp);
3161 }
3162 }
3163 }
3164 }
3165
3166 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
3167 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3168 Node* barrier_node = use->fast_out(i);
3169 if (barrier_set->is_gc_barrier_node(barrier_node)) {
3170 worklist.push(barrier_node);
3171 }
3172 }
3173 }
3174
3175 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3176 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3177 // value changed.
3178 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3179 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3180 uint use_op = use->Opcode();
3181 if (
3182 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3183 (parent_type != nullptr && parent_type->is_con()) ||
3184 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3185 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3186
3187 auto push_and_uses_to_worklist = [&](Node* n) {
3188 uint opc = n->Opcode();
3189 if (opc == Op_AndI || opc == Op_AndL) {
3190 push_if_not_bottom_type(worklist, n);
3191 }
3192 };
3193 auto is_boundary = [](Node* n) {
3194 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3195 };
3196 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3197 }
3198 }
3199
3200 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3201 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3202 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3203 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3204 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3205 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3206 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3207 Node* other_cmp_input = use->in(1);
3208 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
3209 Node* cast_ii = other_cmp_input->fast_out(i);
3210 if (cast_ii->is_CastII()) {
3211 push_if_not_bottom_type(worklist, cast_ii);
3212 }
3213 }
3214 }
3215 }
3216
3217 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3218 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3219 push_if_not_bottom_type(worklist, use->unique_out());
3220 }
3221 }
3222
3223 //------------------------------do_transform-----------------------------------
3224 // Top level driver for the recursive transformer
3225 void PhaseCCP::do_transform() {
3226 // Correct leaves of new-space Nodes; they point to old-space.
3227 C->set_root( transform(C->root())->as_Root() );
3228 assert( C->top(), "missing TOP node" );
3229 assert( C->root(), "missing root" );
3230 }
3231
3232 //------------------------------transform--------------------------------------
3233 // Given a Node in old-space, clone him into new-space.
3234 // Convert any of his old-space children into new-space children.
3235 Node *PhaseCCP::transform( Node *n ) {
3236 assert(n->is_Root(), "traversal must start at root");
3237 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3238
3239 ResourceMark rm;
3240 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3241 Node_List node_map;
3242 // Pre-allocate to avoid frequent realloc
3243 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3244 // track all visited nodes, so that we can remove the complement
3245 Unique_Node_List useful;
3246
3247 if (KillPathsReachableByDeadTypeNode) {
3248 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3249 Node* type_node = _maybe_top_type_nodes.at(i);
3250 if (type(type_node) == Type::TOP) {
3251 ResourceMark rm;
3252 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3253 }
3254 }
3255 } else {
3256 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3257 }
3258
3259 // Initialize the traversal.
3260 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3261 // the logic below doesn't follow any path from Root to the loop body: there's at least one such path but it's proven
3262 // never taken (its type is TOP). As a consequence the node on the exit path that's input to Root (let's call it n) is
3263 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3264 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3265 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3266 // and all safepoints.
3267 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3268 Node* nn = _root_and_safepoints.at(i);
3269 Node* new_node = node_map[nn->_idx];
3270 assert(new_node == nullptr, "");
3271 new_node = transform_once(nn); // Check for constant
3272 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3273 transform_stack.push(new_node); // Process children of cloned node
3274 useful.push(new_node);
3275 }
3276
3277 while (transform_stack.is_nonempty()) {
3278 Node* clone = transform_stack.pop();
3279 uint cnt = clone->req();
3280 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3281 Node *input = clone->in(i);
3282 if( input != nullptr ) { // Ignore nulls
3283 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3284 if( new_input == nullptr ) {
3285 new_input = transform_once(input); // Check for constant
3286 node_map.map( input->_idx, new_input );// Flag as having been cloned
3287 transform_stack.push(new_input); // Process children of cloned node
3288 useful.push(new_input);
3289 }
3290 assert( new_input == clone->in(i), "insanity check");
3291 }
3292 }
3293 }
3294
3295 // The above transformation might lead to subgraphs becoming unreachable from the
3296 // bottom while still being reachable from the top. As a result, nodes in that
3297 // subgraph are not transformed and their bottom types are not updated, leading to
3298 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3299 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3300 // because their address type is inconsistent. Therefore, we aggressively remove
3301 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3302 // to remove them anyway.
3303 if (C->cached_top_node()) {
3304 useful.push(C->cached_top_node());
3305 }
3306 C->update_dead_node_list(useful);
3307 remove_useless_nodes(useful.member_set());
3308 _worklist.remove_useless_nodes(useful.member_set());
3309 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3310
3311 Node* new_root = node_map[n->_idx];
3312 assert(new_root->is_Root(), "transformed root node must be a root node");
3313 return new_root;
3314 }
3315
3316 //------------------------------transform_once---------------------------------
3317 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3318 Node *PhaseCCP::transform_once( Node *n ) {
3319 const Type *t = type(n);
3320 // Constant? Use constant Node instead
3321 if( t->singleton() ) {
3322 Node *nn = n; // Default is to return the original constant
3323 if( t == Type::TOP ) {
3324 // cache my top node on the Compile instance
3325 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3326 C->set_cached_top_node(ConNode::make(Type::TOP));
3327 set_type(C->top(), Type::TOP);
3328 }
3329 nn = C->top();
3330 }
3331 if( !n->is_Con() ) {
3332 if( t != Type::TOP ) {
3333 nn = makecon(t); // ConNode::make(t);
3334 NOT_PRODUCT( inc_constants(); )
3335 } else if( n->is_Region() ) { // Unreachable region
3336 // Note: nn == C->top()
3337 n->set_req(0, nullptr); // Cut selfreference
3338 bool progress = true;
3339 uint max = n->outcnt();
3340 DUIterator i;
3341 while (progress) {
3342 progress = false;
3343 // Eagerly remove dead phis to avoid phis copies creation.
3344 for (i = n->outs(); n->has_out(i); i++) {
3345 Node* m = n->out(i);
3346 if (m->is_Phi()) {
3347 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3348 replace_node(m, nn);
3349 if (max != n->outcnt()) {
3350 progress = true;
3351 i = n->refresh_out_pos(i);
3352 max = n->outcnt();
3353 }
3354 }
3355 }
3356 }
3357 }
3358 replace_node(n,nn); // Update DefUse edges for new constant
3359 }
3360 return nn;
3361 }
3362
3363 // If x is a TypeNode, capture any more-precise type permanently into Node
3364 if (t != n->bottom_type()) {
3365 hash_delete(n); // changing bottom type may force a rehash
3366 n->raise_bottom_type(t);
3367 _worklist.push(n); // n re-enters the hash table via the worklist
3368 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3369 }
3370
3371 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3372 switch( n->Opcode() ) {
3373 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3374 case Op_CallDynamicJava:
3375 case Op_FastLock: // Revisit FastLocks for lock coarsening
3376 case Op_If:
3377 case Op_CountedLoopEnd:
3378 case Op_Region:
3379 case Op_Loop:
3380 case Op_CountedLoop:
3381 case Op_Conv2B:
3382 case Op_Opaque1:
3383 _worklist.push(n);
3384 break;
3385 default:
3386 break;
3387 }
3388
3389 return n;
3390 }
3391
3392 //---------------------------------saturate------------------------------------
3393 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3394 const Type* limit_type) const {
3395 const Type* wide_type = new_type->widen(old_type, limit_type);
3396 if (wide_type != new_type) { // did we widen?
3397 // If so, we may have widened beyond the limit type. Clip it back down.
3398 new_type = wide_type->filter(limit_type);
3399 }
3400 return new_type;
3401 }
3402
3403 //------------------------------print_statistics-------------------------------
3404 #ifndef PRODUCT
3405 void PhaseCCP::print_statistics() {
3406 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3407 }
3408 #endif
3409
3410
3411 //=============================================================================
3412 #ifndef PRODUCT
3413 uint PhasePeephole::_total_peepholes = 0;
3414 #endif
3415 //------------------------------PhasePeephole----------------------------------
3416 // Conditional Constant Propagation, ala Wegman & Zadeck
3417 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3418 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3419 NOT_PRODUCT( clear_peepholes(); )
3420 }
3421
3422 #ifndef PRODUCT
3423 //------------------------------~PhasePeephole---------------------------------
3424 PhasePeephole::~PhasePeephole() {
3425 _total_peepholes += count_peepholes();
3426 }
3427 #endif
3428
3429 //------------------------------transform--------------------------------------
3430 Node *PhasePeephole::transform( Node *n ) {
3431 ShouldNotCallThis();
3432 return nullptr;
3433 }
3434
3435 //------------------------------do_transform-----------------------------------
3436 void PhasePeephole::do_transform() {
3437 bool method_name_not_printed = true;
3438
3439 // Examine each basic block
3440 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3441 Block* block = _cfg.get_block(block_number);
3442 bool block_not_printed = true;
3443
3444 for (bool progress = true; progress;) {
3445 progress = false;
3446 // block->end_idx() not valid after PhaseRegAlloc
3447 uint end_index = block->number_of_nodes();
3448 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3449 Node *n = block->get_node(instruction_index);
3450 if( n->is_Mach() ) {
3451 MachNode *m = n->as_Mach();
3452 // check for peephole opportunities
3453 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3454 if( result != -1 ) {
3455 #ifndef PRODUCT
3456 if( PrintOptoPeephole ) {
3457 // Print method, first time only
3458 if( C->method() && method_name_not_printed ) {
3459 C->method()->print_short_name(); tty->cr();
3460 method_name_not_printed = false;
3461 }
3462 // Print this block
3463 if( Verbose && block_not_printed) {
3464 tty->print_cr("in block");
3465 block->dump();
3466 block_not_printed = false;
3467 }
3468 // Print the peephole number
3469 tty->print_cr("peephole number: %d", result);
3470 }
3471 inc_peepholes();
3472 #endif
3473 // Set progress, start again
3474 progress = true;
3475 break;
3476 }
3477 }
3478 }
3479 }
3480 }
3481 }
3482
3483 //------------------------------print_statistics-------------------------------
3484 #ifndef PRODUCT
3485 void PhasePeephole::print_statistics() {
3486 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3487 }
3488 #endif
3489
3490
3491 //=============================================================================
3492 //------------------------------set_req_X--------------------------------------
3493 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3494 assert( is_not_dead(n), "can not use dead node");
3495 #ifdef ASSERT
3496 if (igvn->hash_find(this) == this) {
3497 tty->print_cr("Need to remove from hash before changing edges");
3498 this->dump(1);
3499 tty->print_cr("Set at i = %d", i);
3500 n->dump();
3501 assert(false, "Need to remove from hash before changing edges");
3502 }
3503 #endif
3504 Node *old = in(i);
3505 set_req(i, n);
3506
3507 // old goes dead?
3508 if( old ) {
3509 switch (old->outcnt()) {
3510 case 0:
3511 // Put into the worklist to kill later. We do not kill it now because the
3512 // recursive kill will delete the current node (this) if dead-loop exists
3513 if (!old->is_top())
3514 igvn->_worklist.push( old );
3515 break;
3516 case 1:
3517 if( old->is_Store() || old->has_special_unique_user() )
3518 igvn->add_users_to_worklist( old );
3519 break;
3520 case 2:
3521 if( old->is_Store() )
3522 igvn->add_users_to_worklist( old );
3523 if( old->Opcode() == Op_Region )
3524 igvn->_worklist.push(old);
3525 break;
3526 case 3:
3527 if( old->Opcode() == Op_Region ) {
3528 igvn->_worklist.push(old);
3529 igvn->add_users_to_worklist( old );
3530 }
3531 break;
3532 default:
3533 break;
3534 }
3535
3536 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
3537 }
3538 }
3539
3540 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3541 PhaseIterGVN* igvn = gvn->is_IterGVN();
3542 if (igvn == nullptr) {
3543 set_req(i, n);
3544 return;
3545 }
3546 set_req_X(i, n, igvn);
3547 }
3548
3549 //-------------------------------replace_by-----------------------------------
3550 // Using def-use info, replace one node for another. Follow the def-use info
3551 // to all users of the OLD node. Then make all uses point to the NEW node.
3552 void Node::replace_by(Node *new_node) {
3553 assert(!is_top(), "top node has no DU info");
3554 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3555 Node* use = last_out(i);
3556 uint uses_found = 0;
3557 for (uint j = 0; j < use->len(); j++) {
3558 if (use->in(j) == this) {
3559 if (j < use->req())
3560 use->set_req(j, new_node);
3561 else use->set_prec(j, new_node);
3562 uses_found++;
3563 }
3564 }
3565 i -= uses_found; // we deleted 1 or more copies of this edge
3566 }
3567 }
3568
3569 //=============================================================================
3570 //-----------------------------------------------------------------------------
3571 void Type_Array::grow( uint i ) {
3572 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3573 if( !_max ) {
3574 _max = 1;
3575 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3576 _types[0] = nullptr;
3577 }
3578 uint old = _max;
3579 _max = next_power_of_2(i);
3580 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3581 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3582 }
3583
3584 //------------------------------dump-------------------------------------------
3585 #ifndef PRODUCT
3586 void Type_Array::dump() const {
3587 uint max = Size();
3588 for( uint i = 0; i < max; i++ ) {
3589 if( _types[i] != nullptr ) {
3590 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3591 }
3592 }
3593 }
3594 #endif